Veille CVE

Recherche « Harness »

49 CVE

Critique 7 Élevée 14 Moyenne 20 Faible 4 Non évaluée 4 Exploitées (KEV) Tri : plus récentes « Harness » ✕
  1. CVE-2026-114362 6.3 Moyenne

    HarnessRouter

    A flaw has been found in HarnessRouter up to 0.31.12. This impacts an unknown function of the file gateway/app.py of the component Targets Test Endpoint. Executing a manipulation of the argument base_url can lead to server-side request forgery. The attack may be performed from remote. Upgrading to version 0.31.13 will fix this issue. This patch is called 1c4b1fc1f89737dabd089ac1d1c174a57f764e28. It is suggested to upgrade the affected component.

  2. CVE-2026-114361 5.6 Moyenne

    HarnessRouter

    A vulnerability was detected in HarnessRouter up to 0.31.10. This affects an unknown function of the file gateway/app.py of the component Identity Mode. Performing a manipulation of the argument X-Harness-Org/X-Harness-Member results in improper authorization. The attack is possible to be carried out remotely. The attack is considered to have high complexity. The exploitability is reported as difficult. Upgrading to version 0.31.11 mitigates this issue. The patch is named 48a26cdf9c4544d2c88d1a07f0777f063f1d753f. The affected component should be upgraded.

  3. CVE-2026-114386 4.8 Moyenne

    pydantic-ai-harness

    Pydantic AI Harness before 0.59.0 fails to enforce UsageLimits.tool_calls_limit on nested tool calls dispatched from run_code in the CodeMode toolset. Users or prompt-injected content can make the sandboxed program invoke tools via run_nested_call() without check_before_tool_call, exceeding operator limits on costly or side-effecting tools.

  4. CVE-2026-17557 6.5 Moyenne

    Red Hat OpenShift AI (RHOAI)

    A flaw was found in the `lm-evaluation-harness` component of Red Hat OpenShift AI. An authenticated user, able to submit an evaluation job, could exploit this by providing a malicious model URL in the `JobSpec`. This vulnerability allows for the disclosure of the tenant's model authentication secret to an attacker-controlled external listener. Additionally, it enables Server-Side Request Forgery (SSRF) attacks against internal network endpoints reachable from the evaluation job pod.

  5. CVE-2026-49410 8.1 Élevée

    tinyhumansai openhuman

    OpenHuman is an open-source agent harness available as a desktop application and Rust library. Versions before 0.56.0 allow cross-site request forgery against the unauthenticated GET /auth/telegram loopback callback because telegram_auth_handler consumes an attacker-supplied one-time Telegram login token and passes the returned JWT to credentials::ops::store_session without validating browser fetch metadata or the request origin. An attacker with an OpenHuman account can cause a user with the local core running to visit a malicious page that requests the /auth/telegram endpoint on loopback port 7788, switching the user's active local session to the attacker's backend account and exposing subsequent agent actions, memory writes, and Telegram messages associated with that session to the attacker. This issue is fixed in version 0.56.0.

  6. CVE-2026-95183 8.8 Élevée

    An issue in the context-compaction feature of DeepSeek Harness (dsh) commit c291e796 allows attackers to execute a prompt injection attack via crafted external content.

  7. CVE-2026-95182 10.0 Critique

    An untrusted search path in DeepSeek Harness (dsh) commit c291e796 allows attackers to execute arbitrary code via supplying and opening crafted executable files in place of legitimate content.

  8. CVE-2026-108789 5.5 Moyenne

    deepseek-ai deepseek-harness

    A vulnerability was identified in deepseek-ai deepseek-harness up to 0.1.1-rc.2. This affects the function agent of the component dsh-workflow-worker-thread. The manipulation leads to enforcement of behavioral workflow. The attack is possible to be carried out remotely. The exploit is publicly available and might be used. The vendor was contacted early about this disclosure but did not respond in any way.

  9. CVE-2026-108788 5.3 Moyenne

    deepseek-ai deepseek-harness

    A vulnerability was determined in deepseek-ai deepseek-harness up to 0.1.1-rc.2. The impacted element is the function spawn of the component Sandbox. Executing a manipulation of the argument Profile can lead to sandbox issue. It is possible to launch the attack on the local host. The exploit has been publicly disclosed and may be utilized. The vendor was contacted early about this disclosure but did not respond in any way.

  10. CVE-2026-71890 8.7 Élevée

    Legion of the Bouncy Castle Inc. BC-JAVA

    In Bouncy Castle for Java before 1.86, validation of an MLS (RFC 9420) external commit's proposal list, org.bouncycastle.mls.protocol.Group.validateExternalCachedProposals, counted the proposals by type and bounded the removed leaf index but never established that the removed leaf had anything to do with the joiner. RFC 9420 sec. 12.2 permits at most one Remove proposal in an external commit, with which the joiner removes an old version of themselves, and requires that where one is present the LeafNode in the commit's path field meet the criteria it would have to meet in an Update for the removed leaf, in particular that its credential present identifiers acceptable for the removed participant. The ordinary proposal-list validator's self-remove rule is deliberately not applied on this path, because a resync commit legitimately removes a leaf the joiner owns, but nothing was put in its place. Any party holding the group's public GroupInfo, which is precisely what an external joiner is meant to be given, could therefore commit a Remove naming any member's LeafIndex and have every member apply it, evicting that member and taking over their slot in the ratchet tree. The credential check that should have prevented this existed only in the gRPC interop harness and so protected no other caller of the public Group.externalJoin and Group.handle API. An external commit carrying a Remove is now accepted only when the removed leaf's credential is identical to the one in the joiner's own new leaf, on both the sending and the receiving side.

  11. CVE-2026-102720 5.3 Moyenne

    Eclipse Foundation eclipse-threadx/netxduo

    A DHCP server, or anyone on the LAN who answers a DISCOVER first, can make the client read about a kilobyte past the end of the received message. The option walk keeps a pointer and an offset in step, and the only bound check uses the offset: ```c /* addons/dhcp/nxd_dhcp_client.c:7538, 7572 */ while (i < length - 1) { ... size = *(++data); /* data moves 1: type -> length byte */ data += size + 1; /* data moves size + 1 more */ i += size + 1; /* i moves only size + 1 */ } ``` A TLV option occupies size + 2 bytes. `data` is advanced by size + 2 in total, `i` by size + 1, so the offset falls one byte behind the real read position for every option the walk skips. After enough skipped options the check `i < length - 1` still holds while `data` is already past the end of the message, and the subsequent read of the type and length bytes comes from whatever follows. A single OFFER carrying a long run of skippable options is enough: ``` ERROR: AddressSanitizer: heap-buffer-overflow READ of size 1 at 0x61b000000794 thread T5 #0 _nx_dhcp_search_buffer addons/dhcp/nxd_dhcp_client.c:7541 #1 _nx_dhcp_get_option_value addons/dhcp/nxd_dhcp_client.c:7082 0x61b000000794 is located 164 bytes to the right of 1648-byte region ``` A well formed OFFER through the same path is handled normally, the client records the offer and moves to REQUESTING, so the difference is the option layout rather than the harness. The read runs in the DHCP client thread while the client is still unconfigured, so it happens on every boot in reach of a hostile DHCP responder. The values read are used to configure the interface, which is how the disclosed bytes become observable. Advance `i` by size + 2, or derive the bound from `data` rather than keeping a second counter.

  12. CVE-2026-101132 3.1 Faible

    deepseek-harness

    A security flaw has been discovered in DeepSeek deepseek-harness up to 0.1.7-rc.2. The affected element is the function loadProfile of the file packages/boot/app-boot/src/profile.ts of the component Bundle Patch Handler. The manipulation of the argument dsh.bundle.patch results in path traversal. The attack may be performed from remote. The attack requires a high level of complexity. The exploitability is described as difficult. The exploit has been released to the public and may be used for attacks. The vendor was contacted early about this disclosure but did not respond in any way.

  13. CVE-2026-101131 3.3 Faible

    deepseek-ai deepseek-harness

    A vulnerability was identified in deepseek-ai deepseek-harness up to 0.1.5-rc.3. Impacted is an unknown function of the file packages/e2b/e2b/src/index.ts of the component dsh. The manipulation of the argument E2B_API_KEY leads to reliance on untrusted inputs in a security decision. Local access is required to approach this attack. The exploit is publicly available and might be used. The vendor was contacted early about this disclosure but did not respond in any way.

  14. CVE-2026-101102 6.3 Moyenne

    deepseek-ai deepseek-harness

    A vulnerability was found in deepseek-ai deepseek-harness up to 0.1.0-rc.7. Impacted is the function run_code of the component Code Mode Sandbox. The manipulation results in sandbox issue. The attack can be executed remotely. The vendor's own code, SAFETY.md, and design notes all explicitly state the worker is "containment, not a security boundary". The vendor was contacted early about this disclosure but did not respond in any way.

  15. CVE-2026-101078 6.3 Moyenne

    deepseek-ai deepseek-harness

    A vulnerability has been found in deepseek-ai deepseek-harness up to 0.1.7-rc.2. Affected is an unknown function of the file packages/sandbox/sandbox-local/src/profiles.ts of the component Landlock Backend. Such manipulation leads to improper isolation or compartmentalization. The attack must be carried out locally. The exploit has been disclosed to the public and may be used. It is advisable to implement a patch to correct this issue. The vendor was contacted early about this disclosure but did not respond in any way.

  16. CVE-2026-100601 5.3 Moyenne

    openclaw clawhub

    ClawHub (openclaw/clawhub) application/backend contains a server-side request forgery vulnerability in the public profile preview's image fetching. The preview accepts a user-supplied image URL and checks the textual hostname against private-address patterns, but does not validate or pin the resolved network destination, so a public-looking hostname can resolve to an internal address or change resolution between validation and connection (DNS rebinding). A maintainer-run local harness demonstrated an outbound connection to an owner-controlled loopback listener; access to production internal services, credential disclosure, and code execution were not demonstrated. The issue was confirmed at revision cbfee7343ddc867316dd9b3de6fa8856730f9f41; the complete historical affected range was not established. Fixed by PR #3683, included in revision 8c2de6c506bb4efabe3f0c2ffb8370b9e23d4650, which was deployed to clawhub.ai on 2026-09-11; self-hosted deployments should update to that revision or a later descendant. The npm CLI and OpenClaw runtime are separate products and are not affected.

  17. CVE-2026-54918 5.3 Moyenne

    netbox-community devicetype-library

    NetBox Device Type Library is a collection of community-sourced device type definitions for import into NetBox. In the affected repository revisions, NETBOX_DT_LIBRARY_URL in tests/test_configuration.py is a free-form tracked constant that an unauthenticated pull-request author can change before the validation test harness runs. During pytest collection, tests/definitions_test.py passes the value to Repo.clone_from and create_remote("upstream").fetch(), causing blind Git smart-HTTP requests to an attacker-selected host or loading attacker-controlled tests/known-*.json validation caches. The blind request cannot set arbitrary metadata-service headers or return response bodies, and this path does not execute remote Git hooks, but substituted known data can bypass slug, module, and rack uniqueness validation. This vulnerability is fixed in commit 8980c690097e92f5028c7e6df402b327d827ecd5.

  18. CVE-2026-54752 9.6 Critique

    netbox-community devicetype-library

    NetBox Device Type Library is a collection of community-sourced device type definitions for import into NetBox. The validation test harness can deserialize pull-request-controlled tracked pickle cache files through pickle.load in the read_pickle_data function in tests/pickle_operations.py. An unauthenticated contributor can change USE_LOCAL_KNOWN_SLUGS in tests/test_configuration.py and supply a crafted tests/known-modules.pickle or tests/known-racks.pickle file that tests/definitions_test.py loads when pytest runs. Deserialization invokes attacker-controlled object reduction behavior, allowing arbitrary code execution in the GitHub Actions runner or in a maintainer process that runs the tests, with the confidentiality, integrity, and availability of reachable resources at risk. This vulnerability is fixed with commit 1c6f7e2b93589b965318c6e67ac3504831f0e71e.

  19. CVE-2026-90153 8.1 Élevée

    Linux

    In the Linux kernel, the following vulnerability has been resolved: ksmbd: bound smb_check_perm_dacl() ACE walks by DACL size smb_check_perm_dacl() validates that the DACL fits inside the NT security descriptor, but then bounds its two ACE walks by the remaining NTSD length (acl_size) rather than the DACL's declared size (pdacl_size). When pdacl->size is smaller than the trailing NTSD buffer, bytes after the declared DACL boundary - still inside the stored security descriptor - are parsed as ACEs during access checks. A crafted DACL can place an access-granting ACE beyond pdacl->size, and the current code accepts it during SMB2_CREATE access validation, while parse_dacl() and smb_inherit_dacl() stop at pdacl_size. Bound both ACE walks by pdacl_size to match the DACL boundary semantics used elsewhere in the server. Validation: - semantic KUnit harness shows the post-boundary ACE is selected before the fix and rejected (EACCES) after it - linux master (7.2-rc6), x86_64

  20. CVE-2026-92750 6.5 Moyenne

    harness

    Harness through 3.3.0 omits access control validation in the infrastructure provider read endpoint, allowing authenticated users to retrieve provider configurations from spaces they do not belong to. Attackers can query the GET /api/v1/infraproviders endpoint with arbitrary space identifiers to expose sensitive provider metadata including Docker endpoints, TLS certificate paths, and cloud project identifiers.

  21. CVE-2026-82533 9.6 Critique

    DeepSeek Harness

    DeepSeek Harness before 0.1.2-alpha.1 contains an authentication bypass vulnerability that grants unauthenticated access to its local HTTP agent-control API by accepting a client-supplied loopback Host header in place of validating the actual TCP connection origin. On the default configuration, a confined tool-executed process can reach the loopback API without any port exposure and use it to escape its own OS sandbox, escalate to unconfined execution, and disable the approval prompt. When the port is externally reachable via tunnel, SSH forward, or reverse proxy, a remote attacker can exploit the same flaw to create sessions, execute arbitrary commands, and exfiltrate stored conversation transcripts without credentials.

  22. CVE-2026-80808 – Non évaluée

    Linux

    In the Linux kernel, the following vulnerability has been resolved: ext4: stop retrying saturated xattr cache entries ext4_xattr_block_set() retries when a cache entry selected for reuse has a saturated reference count after taking the buffer lock. The retry returns to the mbcache lookup without making that entry ineligible, so it can select the same unusable entry indefinitely. A task spinning there can hold the parent directory's i_rwsem and leave concurrent rmdir callers blocked. Normally a reusable entry has a reference count below EXT4_XATTR_REFCOUNT_MAX because the count and MBE_REUSABLE_B are updated under the same buffer lock. A corrupted filesystem can violate that invariant. The syzbot reproducer reports allocator and xattr corruption before triggering this retry loop. Check the untrusted on-disk count before incrementing it, avoiding overflow, and clear MBE_REUSABLE_B when it is already saturated. The next lookup then skips the entry that was just proven unusable. This mirrors the normal transition at EXT4_XATTR_REFCOUNT_MAX; the release path marks the entry reusable again on the exact 1024-to-1023 transition. Using the same QEMU harness and guest parameters, current unpatched Linux hung in 6 of 8 420-second trials with the do_rmdir signature; representative NMI backtraces caught the owner spinning in ext4_xattr_block_set(). The patched kernel completed 28 of 28 trials without a hung-task report; the final twelve trials exercised the reviewed overflow-safe form of the change. syzbot's patch testing also completed without reproducing the hang.

  23. CVE-2026-80784 – Non évaluée

    Linux

    In the Linux kernel, the following vulnerability has been resolved: mptcp: pm: fix memory leak from alloc-during-teardown race mptcp_pm_destroy() empties msk->pm.anno_list and msk->pm.userspace_pm_local_addr_list under msk->pm.lock during socket teardown, dropping the lock between the two. A concurrent userspace PM genl ANNOUNCE on the same msk holds a sock reference via mptcp_token_get_sock() and, in mptcp_pm_nl_announce_doit(), calls mptcp_userspace_pm_append_new_local_addr() and mptcp_pm_announced_alloc(). Both take msk->pm.lock briefly to add to their respective lists. Because the genl handler holds a sock reference, mptcp_pm_destroy() may run on the same msk via mptcp_disconnect(), which invokes mptcp_destroy_common() without dropping the sock refcount, before the handler completes. If the lock acquisitions interleave such that mptcp_pm_destroy() empties a list first, the later alloc adds its entry to a list head that nothing else iterates for this msk, and the entry leaks. kmemleak reports both mptcp_pm_add_addr objects (from mptcp_pm_announced_alloc()) and mptcp_pm_addr_entry objects (from mptcp_userspace_pm_append_new_local_addr()) under sustained concurrent ANNOUNCE + close load against the userspace PM. Add an MPTCP_PM_DESTROYING bit in msk->pm.status, set by mptcp_pm_destroy() under pm.lock before the lists are emptied and checked under pm.lock by the alloc paths. Either the alloc takes pm.lock first, in which case its entry is on the list when mptcp_pm_destroy() frees it; or mptcp_pm_destroy() takes pm.lock first, in which case the later alloc observes the bit and refuses. Found by an MPTCP protocol-flow harness extending BRF (arXiv:2305.08782).

  24. CVE-2026-62677 8.8 Élevée

    omnigent-ai omnigent

    Omnigent is an open-source AI agent framework and meta-harness for orchestrating coding agents. Prior to 0.3.0, an authenticated user can upload a session-scoped agent bundle with an absolute or traversal-containing os_env.cwd value because omnigent/spec/parser.py stores the value verbatim and omnigent/spec/validator.py does not constrain it. On a runner where OMNIGENT_RUNNER_WORKSPACE is unset, omnigent/runner/resource_registry.py preserves the attacker-controlled path and omnigent/inner/os_env.py uses the resolved path as the environment root and copytree source. The _assert_within_cwd check then treats that attacker-selected root as trusted, allowing sys_os_read, write, edit, and shell tools to access runner files and environment secrets outside the intended workspace. This issue is fixed in version 0.3.0.

  25. CVE-2026-62676 7.1 Élevée

    omnigent-ai omnigent

    Omnigent is an open-source AI agent framework and meta-harness for orchestrating coding agents. Prior to 0.3.0, the shared shell-command parser in omnigent/policies/builtins/_shell.py fails to recognize combined interpreter flags, the timeout, nice, setsid, and stdbuf wrappers, command substitutions, and a single background control operator. A gated git push or gh write hidden with these forms produces no parsed operation, causing the github.py write_repos and write_branches allowlist and the working_dir.py workspace confinement policies to abstain and allow the command. An authenticated or prompt-injected agent can therefore push to an unauthorized repository or branch or escape the intended workspace. This issue is fixed in version 0.3.0.

  26. CVE-2026-62675 8.8 Élevée

    omnigent-ai omnigent

    Omnigent is an open-source AI agent framework and meta-harness for orchestrating coding agents. Prior to 0.3.0, multipart POST /v1/sessions accepts an authenticated user's agent bundle and omnigent/server/bundles.py validate_agent_bundle does not reject a tools..callable dotted Python path. omnigent/runner/tool_dispatch.py _resolve_spec_callable imports the specified module and _execute_spec_callable_tool invokes the resolved function, allowing a bundle to select subprocess.check_output and execute a local command with the runner process permissions. This can expose runner files, environment variables, credentials, workspace data, internal services, and availability without administrator access. This issue is fixed in version 0.3.0.

  27. CVE-2026-62674 9.0 Critique

    omnigent-ai omnigent

    Omnigent is an open-source AI agent framework and meta-harness for orchestrating coding agents. Prior to 0.3.0, PUT /sessions/{session_id}/agent checks LEVEL_EDIT permission for a session but does not reject a bound shared or template agent whose agent.session_id is None. An authenticated user with edit access to a session can replace that shared agent bundle through omnigent/server/routes/sessions.py, add a stdio MCP server, and cause later sessions that use the shared agent to launch an attacker-controlled command through omnigent/tools/mcp.py. The command executes with the Omnigent runner process permissions and can expose files, credentials, workspace data, internal services, and runner availability. This issue is fixed in version 0.3.0.

  28. CVE-2026-72193 – Non évaluée

    Linux

    In the Linux kernel, the following vulnerability has been resolved: ntfs3: cap RESTART_TABLE free-chain walker at rt->used A crafted NTFS3 disk image triggers an in-kernel infinite loop at mount time, hanging the mounting thread and firing the soft-lockup watchdog within ~22s on multi-CPU hosts (panic with kernel.softlockup_panic=1). The bug is reachable from desktop USB auto-mount on distributions where udisks2 routes the NTFS signature to the in-tree ntfs3 driver (Arch family and an increasing fraction of Fedora / openSUSE / RHEL deployments); CAP_SYS_ADMIN-class manual mount elsewhere. check_rstbl()'s second walker iterates the free-entry singly-linked list headed by rt->first_free with no upper bound on iteration count: for (off = ff; off;) { if (off == RESTART_ENTRY_ALLOCATED) return false; off = le32_to_cpu(*(__le32 *)Add2Ptr(rt, off)); if (off > ts - sizeof(__le32)) return false; } The existing guards cover three exits: end-of-list (off == 0), the in-use marker (off == RESTART_ENTRY_ALLOCATED), and out-of-bounds (off > ts - sizeof(__le32)). None of the three prevents an in-bounds cycle. A crafted on-disk RESTART_TABLE whose free chain contains a self-loop or A->B->A cycle whose offsets satisfy: - in range [sizeof(struct RESTART_TABLE), ts - sizeof(__le32)] - (off - sizeof(struct RESTART_TABLE)) % rsize == 0 passes all existing guards and spins the mount-time thread forever. Reproduced in UML by hand-forging a 2 MB NTFS3 image whose journal RESTART_TABLE first_free = 0x18 and whose entry at offset 0x18 stores 0x18 as its next pointer; mount of the forged image with the in-tree ntfs3 driver never returns. Bound the walker by rt->used. Each entry on a legitimate free chain is unique, and the total slot count is ne = le16_to_cpu (rt->used). A traversal that visits more than ne slots is by construction malformed; reject it as a corrupt RESTART_TABLE. After this patch, mount of the forged image returns with -EINVAL and a log_replay failure message, and mkntfs-produced legitimate images mount cleanly (verified in the same UML harness).

  29. CVE-2026-18830 8.1 Élevée

    AWS Amazon Bedrock AgentCore harness

    Insufficient input validation in Amazon Bedrock AgentCore harness might allow an authenticated remote user to execute configured tools bypassing model invocation and security controls via crafted content blocks in conversation messages. AWS has addressed this issue. No customer action is required.

  30. CVE-2026-64456 7.7 Élevée

    Linux

    In the Linux kernel, the following vulnerability has been resolved: hwrng: virtio: clamp device-reported used.len at copy_data() random_recv_done() stores the device-reported used.len directly into vi->data_avail. copy_data() then indexes vi->data[] using vi->data_idx (advanced by previous copy_data() calls) and issues a memcpy() without re-validating either value against the posted buffer size sizeof(vi->data) (SMP_CACHE_BYTES bytes, typically 32 or 64). A malicious or buggy virtio-rng backend can set used.len beyond sizeof(vi->data), steering the memcpy() past the end of the inline array into adjacent kmalloc-1k slab bytes. hwrng_fillfn() mixes those bytes into the guest RNG, and guest root can also observe them directly via /dev/hwrng. Concrete impact is inside the guest: - Memory-safety / hardening: any virtio-rng backend that over-reports used.len causes the driver to read past vi->data into unrelated slab contents. hwrng_fillfn() is a kernel thread that runs as soon as the device is probed; no guest userspace interaction is required to first-trigger the OOB. - Cross-boundary leak (confidential-compute threat model): a malicious hypervisor cooperating with a malicious or compromised guest root userspace can use /dev/hwrng as a leak channel for guest-kernel heap data. The host sets a large used.len, guest root reads /dev/hwrng, and the returned bytes contain guest kernel slab contents that were adjacent to vi->data. In practice, confidential-compute guests (SEV-SNP, TDX) usually disable virtio-rng entirely, so this path is narrow, but the fix is still worth carrying because the underlying memory-safety bug contaminates the guest RNG on any host. KASAN confirms the OOB on a 7.1-rc4 guest whose virtio-rng backend has been patched to report used.len = 0x10000: BUG: KASAN: slab-out-of-bounds in virtio_read+0x394/0x5d0 Read of size 64 at addr ffff88800ae0ba20 by task hwrng/52 Call Trace: __asan_memcpy+0x23/0x60 virtio_read+0x394/0x5d0 hwrng_fillfn+0xb2/0x470 kthread+0x2cc/0x3a0 Allocated by task 1: probe_common+0xa5/0x660 virtio_dev_probe+0x549/0xbc0 The buggy address belongs to the object at ffff88800ae0b800 which belongs to the cache kmalloc-1k of size 1024 The buggy address is located 0 bytes to the right of allocated 544-byte region [ffff88800ae0b800, ffff88800ae0ba20) Same class of bug as commit c04db81cd028 ("net/9p: Fix buffer overflow in USB transport layer"), which hardened usb9pfs_rx_complete() against unchecked device-reported length in the USB 9p transport. With the clamp at point of use and array_index_nospec() in place, the same harness boots cleanly: copy_data() returns zero for the bogus report, the device-supplied bytes after data_idx are discarded, and the driver issues a fresh request.

  31. CVE-2026-64651 6.3 Moyenne

    vercel @ai-sdk/harness-opencode

    The `@ai-sdk/harness-opencode` tool connects HarnessAgent to OpenCode through a sandboxed bridge. Prior to version 1.0.28, the tool relay authorizes requests from any process whose command line contains an allowed helper script path (`host-tool-mcp.mjs`). This allows untrusted code executing in the sandbox to invoke arbitrary host-exposed tools including secret lookups, deployment operations, and cloud API calls without a corresponding model-authorized tool-call event. Exploitation requires a Linux environment (the vulnerable fallback checks `process.platform === 'linux'` and reads `/proc`); an active harness session with one or more host-provided tools; and untrusted code executing in the sandbox (e.g. a malicious dependency, build script, or lifecycle hook) The fix in version 1.0.28 removes the process-path authorization fallback entirely. Relay requests are now only accepted after exact, short-lived, one-time authorization matching the tool name and input from a bridge-observed model event. Some workarounds are available. Do not run the OpenCode harness on untrusted repositories or with untrusted dependencies, and/or limit host-exposed tools to non-sensitive operations when working with untrusted code.

  32. CVE-2026-64650 6.3 Moyenne

    vercel @ai-sdk/harness-codex

    The `@ai-sdk/harness-opencode` tool is an HarnessV1 adapter backed by @openai/codex-sdk, which drives the codex command line interface. Prior to version 1.0.29, the tool relay authorizes requests from any process whose command line contains an allowed helper script path (the Codex CLI shim). This allows untrusted code executing in the sandbox to invoke arbitrary host-exposed tools, including secret lookups, deployment operations, and cloud API calls without a corresponding model-authorized tool-call event. Exploitation requires a Linux environment (the vulnerable fallback checks `process.platform === 'linux'` and reads `/proc`); an active harness session with one or more host-provided tools; and untrusted code executing in the sandbox (e.g. a malicious dependency, build script, or lifecycle hook) The fix in version 1.0.29 removes the process-path authorization fallback entirely. Relay requests are now only accepted after exact, short-lived, one-time authorization matching the tool name and input from a bridge-observed model event. Some workarounds are available. Do not run the Codex harness on untrusted repositories or with untrusted dependencies, and/or limit host-exposed tools to non-sensitive operations when working with untrusted code.

  33. CVE-2026-64102 9.8 Critique

    Linux · Linux linux kernel

    In the Linux kernel, the following vulnerability has been resolved: RDMA/siw: Reject MPA FPDU length underflow before signed receive math A malicious connected siw peer can send an iWARP FPDU whose MPA length field (c_hdr->mpa_len, 16 bit big-endian, peer-controlled) is smaller than the fixed DDP/RDMAP header for the announced opcode. Soft-iWARP parses the full header in siw_get_hdr() based on iwarp_pktinfo[opcode] .hdr_len, but never compares mpa_len against that header length. siw_tcp_rx_data() then derives srx->fpdu_part_rem = be16_to_cpu(mpa_len) - fpdu_part_rcvd + MPA_HDR_SIZE; where fpdu_part_rcvd equals iwarp_pktinfo[opcode].hdr_len at this point. For a tagged WRITE (hdr_len 16, MPA_HDR_SIZE 2) the smallest on-wire mpa_len of 0 yields fpdu_part_rem = -14, and any mpa_len below hdr_len - MPA_HDR_SIZE underflows to a negative int. The signed value then flows into siw_proc_write()/siw_proc_rresp() as bytes = min(srx->fpdu_part_rem, srx->skb_new); is handed to siw_check_mem() as an int len (whose interval check addr + len > mem->va + mem->len is satisfied for a valid base when len is negative), and reaches siw_rx_data() -> siw_rx_kva() / siw_rx_umem() -> skb_copy_bits() as a signed copy length. The header copy branch in skb_copy_bits() promotes that to size_t, producing a multi-gigabyte read. KASAN under a KUnit harness that drives the real kernel TCP receive path -- a loopback AF_INET socketpair, the malformed FPDU written via kernel_sendmsg, sk_data_ready firing in softirq, tcp_read_sock dispatching to siw_tcp_rx_data -- reports: BUG: KASAN: use-after-free in skb_copy_bits+0x284/0x480 Read of size 4294967295 at addr ffff888... Call Trace: skb_copy_bits siw_rx_kva siw_rx_data siw_check_mem siw_proc_write siw_tcp_rx_data __tcp_read_sock siw_qp_llp_data_ready tcp_data_ready tcp_data_queue Add the missing invariant at the earliest point where the peer header is fully assembled. iwarp_pktinfo[*].hdr_len - MPA_HDR_SIZE is exactly the value the siw transmitter uses as the minimum mpa_len for each opcode (drivers/infiniband/sw/siw/siw_qp.c:33), so this matches the protocol contract. Out-of-range FPDUs terminate the connection with TERM_ERROR_LAYER_LLP / LLP_ETYPE_MPA / LLP_ECODE_FPDU_START -- which is RFC 5044 Section 8 error code 3 ("Marker and ULPDU Length fields do not agree on the start of an FPDU"), the correct framing-error class for this inconsistency.

  34. CVE-2026-63905 – Non évaluée

    Linux

    In the Linux kernel, the following vulnerability has been resolved: usbip: vudc: Fix use after free bug in vudc_remove due to race condition This patch follows up Zheng Wang's 2023 report of a use-after-free in vudc_remove(). The original thread stalled on Shuah Khan's request for runtime testing of the unplug/unbind path. This patch supplies that testing and keeps Zheng's original fix shape. In vudc_probe(), v_init_timer() binds udc->tr_timer.timer to v_timer(). usbip_sockfd_store() starts the timer via v_start_timer()/v_kick_timer(). vudc_remove() can then free the containing struct vudc while the timer is still pending or executing. KASAN confirms the race on an unpatched x86_64 QEMU guest with CONFIG_KASAN=y, CONFIG_USBIP_VUDC=y, CONFIG_USB_ZERO=y, and a tight loop that repeatedly writes a socket fd to usbip_sockfd, closes the socket pair, and unbinds/rebinds usbip-vudc.0: BUG: KASAN: slab-use-after-free in __run_timer_base.part.0+0x8ba/0x8e0 Write of size 8 at addr ffff888001b80740 by task trigger_and_unb/239 Allocated by task 239: vudc_probe+0x4d/0xaa0 Freed by task 239: kfree+0x18f/0x520 device_release_driver_internal+0x388/0x540 unbind_store+0xd9/0x100 This lands in the timer core rather than v_timer() itself because the embedded timer_list is being walked after its containing struct vudc has already been freed. The underlying lifetime bug is the same one Zheng reported. With v_stop_timer() called from vudc_remove() and the timer deleted synchronously, the same harness completed 5000 bind/unbind iterations with no KASAN report.

  35. CVE-2026-63808 9.8 Critique

    Linux · Linux linux kernel

    In the Linux kernel, the following vulnerability has been resolved: exfat: fix potential use-after-free in exfat_find_dir_entry() In exfat_find_dir_entry(), the buffer_head obtained from exfat_get_dentry() is released with brelse(bh) before the fall-through TYPE_EXTEND branch reads the directory entry through ep (which points into bh->b_data): brelse(bh); if (entry_type == TYPE_EXTEND) { ... len = exfat_extract_uni_name(ep, entry_uniname); ... } After brelse() drops our reference, nothing guarantees that the underlying page backing bh->b_data remains valid for the subsequent exfat_extract_uni_name() read. This is the same pattern fixed in commit fc961522ddbd ("exfat: Fix potential use after free in exfat_load_upcase_table()"). Move brelse(bh) so it runs after ep is no longer dereferenced on each branch. Confirmed on QEMU x86_64 with CONFIG_KASAN=y + CONFIG_DEBUG_PAGEALLOC=y + CONFIG_PAGE_POISONING=y on linux-next, using a crafted exFAT image (long filename with same-hash collisions forcing the TYPE_EXTEND path). With a debug-only invalidate_bdev() inserted between brelse(bh) and the ep read to make the stale-deref window deterministic, the unpatched kernel faults: BUG: KASAN: use-after-free in exfat_find_dir_entry+0x133b/0x15a0 BUG: unable to handle page fault for address: ffff88801a5fa0c2 Oops: 0000 [#1] SMP DEBUG_PAGEALLOC KASAN NOPTI RIP: 0010:exfat_find_dir_entry+0x1188/0x15a0 With this patch applied, the same instrumented harness completes cleanly under the same sanitizer stack. I have not reproduced a crash on an uninstrumented kernel under ordinary reclaim; the instrumented A/B establishes the lifetime violation and that the patch closes it, not an unaided triggerability claim.

  36. CVE-2026-59726 10.0 Critique

    ruvnet ruflo

    Ruflo is an agent meta-harness for Claude Code and Codex. Prior to 3.16.3, ruflo's default docker-compose deployment exposed the MCP bridge POST /mcp and POST /mcp/:group endpoints without authentication, allowing an unauthenticated network attacker to invoke tools/call to terminal_execute, obtain a shell in the bridge container, read provider API keys, and poison AgentDB learning-store patterns. This issue is fixed in version 3.16.3.

  37. CVE-2026-15036 4.3 Moyenne

    Harness

    A vulnerability was determined in Harness up to 2.28.2. This vulnerability affects the function getAuthorizedSpaces of the file app/api/controller/gitspace/list_all.go of the component gitspaces Endpoint. Executing a manipulation can lead to authorization bypass. The attack can be executed remotely. The exploit has been publicly disclosed and may be utilized. The project was informed of the problem early through an issue report but has not responded yet.

  38. CVE-2026-7838 8.8 Élevée

    uvnc UltraVNC

    UltraVNC viewer through 1.8.2.2 contains an integer overflow leading to a heap buffer overflow in the RFB protocol failure-response parsing path. In vncviewer/ClientConnection.cpp, the 4-byte network-supplied reasonLen field (type CARD32) is passed as reasonLen+1 to CheckBufferSize(). Because both operands are unsigned 32-bit, a reasonLen of 0xFFFFFFFF overflows to 0, causing CheckBufferSize to allocate only 256 bytes. The subsequent ReadString(m_netbuf, reasonLen) call then performs ReadExact for the original 4 GiB length into that 256-byte heap buffer. This overflow is reachable via rfbConnFailed (auth-scheme negotiation) and rfbVncAuthFailed (post-handshake) message types without successful authentication. A malicious VNC server, or any man-in-the-middle on the RFB stream, can trigger this condition when the victim viewer connects, potentially resulting in remote code execution as the user running the viewer. The crash was confirmed with AddressSanitizer on a portable reproduction harness (heap-buffer-overflow WRITE at offset 256).

  39. CVE-2026-54328 7.3 Élevée

    earendil-works pi

    Pi is a minimal terminal coding harness. From 0.74.0 until 0.78.1, Pi versions with temporary npm or git extension package installs used predictable paths under the operating system temporary directory. On Linux-based multi-user systems, a local attacker who can write to the shared temporary directory could prepare the expected package location before another user runs pi with a temporary extension package source. Pi could then load attacker-controlled extension code in the victim user's process. This vulnerability is fixed in 0.78.1.

  40. CVE-2026-54327 2.2 Faible

    earendil-works pi

    Pi is a minimal terminal coding harness. From 0.74.0 until 0.78.1, Pi stored API keys and OAuth credentials in auth.json. A race condition in the file write path could briefly create or rewrite this file with permissions derived from the process umask before tightening the file to owner-only permissions. This vulnerability is fixed in 0.78.1.

  41. CVE-2026-54326 2.5 Faible

    earendil-works pi

    Pi is a minimal terminal coding harness. From 0.74.0 until 0.78.1, Pi HTML exports render session Markdown into a static HTML file. It did not consistently reject unsafe Markdown link and image URL schemes. In versions with scheme filtering, C0 control characters in the URL scheme could bypass the check because browsers normalize those characters before navigation. This vulnerability is fixed in 0.78.1.

  42. CVE-2026-54325 4.4 Moyenne

    earendil-works pi

    Pi is a minimal terminal coding harness. Pi before 0.79.0 loaded project-local configuration and resources from a repository's .pi directory without first asking the user to trust that repository. This included project-local extensions, which are executable TypeScript or JavaScript modules loaded into the Pi process. An attacker who controls a repository could place Pi-specific project resources in that repository. If a user then started Pi from that working tree, the project-local extension code could run with the same privileges as the local Pi process without the user having a convenient way to make a trust decision. This vulnerability is fixed in 0.79.0.

  43. CVE-2026-56696 5.4 Moyenne

    HKUDS OpenHarness

    OpenHarness /issue and /pr_comments slash commands lack remote_invocable=False protection, allowing remote channel senders to write attacker-controlled Markdown into project context files. Admitted remote attackers can inject malicious content into .openharness/issue.md and .openharness/pr_comments.md files, which are subsequently injected into runtime system prompts, persistently influencing local agent behavior.

  44. CVE-2026-56695 6.5 Moyenne

    HKUDS OpenHarness

    OpenHarness ohmo gateway /resume and /summary slash commands default remote_invocable to True, allowing admitted remote senders to enumerate and load arbitrary session snapshots by ID. Attackers can exploit this to access victim snapshots containing private prompts, credentials, tool output, and file paths via shared gateway channels.

  45. CVE-2026-49325 4.6 Moyenne

    Indian Motorcycle Scout Bobber + Tech

    Improper handling of physical conditions in the bike-shutdown control of the Indian Motorcycle Scout Bobber + Tech 2025 model year allows a physical attacker with access to the Wireless Control Module (WCM) wiring harness to bypass the anti-theft shutdown. The WCM signals shutdown to a peer ECU via a falling-edge voltage transition on a dedicated wire pair. The receiving ECU does not distinguish between an active shutdown pulse and an open-circuit / disconnected condition; interrupting the relevant wires leaves the motorcycle fully operable even though the WCM never validated the rider's PIN. Specific connector details have been withheld pending vendor remediation.

  46. CVE-2026-46116 7.8 Élevée

    Linux · Red Hat Enterprise Linux 10 · Red Hat Enterprise Linux 10.0 Extended Update Support · Red Hat Enterprise Linux 7 Extended Lifecycle Support

    In the Linux kernel, the following vulnerability has been resolved: xfrm: defensively unhash xfrm_state lists in __xfrm_state_delete KASAN reproduces a slab-use-after-free in __xfrm_state_delete()'s hlist_del_rcu calls under syzkaller load on linux-6.12.y stable (reproduced on 6.12.47, also reachable via the same code path on torvalds/master and on the ipsec tree). Nine unique signatures cluster in the xfrm_state lifecycle, the load-bearing one being: BUG: KASAN: slab-use-after-free in __hlist_del include/linux/list.h:990 [inline] BUG: KASAN: slab-use-after-free in hlist_del_rcu include/linux/rculist.h:516 [inline] BUG: KASAN: slab-use-after-free in __xfrm_state_delete net/xfrm/xfrm_state.c Write of size 8 at addr ffff8881198bcb70 by task kworker/u8:9/435 Workqueue: netns cleanup_net Call Trace: __hlist_del / hlist_del_rcu __xfrm_state_delete xfrm_state_delete xfrm_state_flush xfrm_state_fini ops_exit_list cleanup_net The other observed signatures hit the same slab object from __xfrm_state_lookup, xfrm_alloc_spi, __xfrm_state_insert and an OOB write variant of __xfrm_state_delete, all on the byseq/byspi hash chains. __xfrm_state_delete() guards its byseq and byspi unhashes with value-based predicates: if (x->km.seq) hlist_del_rcu(&x->byseq); if (x->id.spi) hlist_del_rcu(&x->byspi); while everywhere else in the file (e.g. state_cache, state_cache_input) the safer hlist_unhashed() check is used. xfrm_alloc_spi() sets x->id.spi = newspi inside xfrm_state_lock and then immediately inserts into byspi, but a path that observes x->id.spi != 0 outside of xfrm_state_lock can still skip-or-hit the byspi unhash inconsistently with whether x is actually on the list. The same holds for x->km.seq versus byseq, and the bydst/bysrc unhashes have no predicate at all, so a second __xfrm_state_delete() on the same object writes through LIST_POISON pprev. The defensive change here: - Use hlist_del_init_rcu() instead of hlist_del_rcu() on bydst, bysrc, byseq and byspi so a second deletion is a no-op rather than a write through LIST_POISON pprev. The byseq/byspi nodes are already initialised in xfrm_state_alloc(). - Test hlist_unhashed() rather than the value predicate for byseq/byspi, so the unhash decision tracks list state rather than mutable scalar fields. Empirical verification: applied this patch on top of v6.12.47, rebuilt, and re-ran the same syzkaller harness for 1h16m on a previously-crashy configuration that produced ~100 hits each of slab-use-after-free Read in xfrm_alloc_spi / Read in __xfrm_state_lookup / Write in __xfrm_state_delete. After the patch, 7.1M execs across 32 VMs at ~1550 exec/sec produced zero xfrm_state UAF/OOB hits. /proc/slabinfo confirms the xfrm_state slab is actively allocated and freed during the run (~143 KiB resident), so the fuzzer is still exercising those code paths -- they just no longer crash. Reproduction: - Linux 6.12.47 x86_64 + KASAN_GENERIC + KASAN_INLINE + KCOV - syzkaller @ 746545b8b1e4c3a128db8652b340d3df90ce61db - 32 QEMU/KVM VMs x 2 vCPU on AWS c5.metal bare metal - 9 unique signatures collected in ~9h, all within xfrm_state lifecycle

  47. CVE-2026-46072 5.5 Moyenne

    Linux · Linux linux kernel

    In the Linux kernel, the following vulnerability has been resolved: ntfs3: add buffer boundary checks to run_unpack() run_unpack() checks `run_buf < run_last` at the top of the while loop but then reads size_size and offset_size bytes via run_unpack_s64() without verifying they fit within the remaining buffer. A crafted NTFS image with truncated run data in an MFT attribute triggers an OOB heap read of up to 15 bytes when the filesystem is mounted. Add boundary checks before each run_unpack_s64() call to ensure the declared field size does not exceed the remaining buffer. Found by fuzzing with a source-patched harness (LibAFL + QEMU).

  48. CVE-2026-46062 7.8 Élevée

    Linux · Linux linux kernel

    In the Linux kernel, the following vulnerability has been resolved: ntfs3: fix integer overflow in run_unpack() volume boundary check The volume boundary check `lcn + len > sbi->used.bitmap.nbits` uses raw addition which can wrap around for large lcn and len values, bypassing the validation. Use check_add_overflow() as is already done for the adjacent prev_lcn + dlcn and vcn64 + len checks added by commit 3ac37e100385 ("ntfs3: Fix integer overflow in run_unpack()"). Found by fuzzing with a source-patched harness (LibAFL + QEMU).

  49. CVE-2026-7551 8.8 Élevée

    HKUDS OpenHarness

    HKUDS OpenHarness contains a remote code execution vulnerability in the /bridge slash command that allows remote senders accepted by configuration to execute arbitrary operating system commands. Attackers can invoke the /bridge spawn command with attacker-controlled command text that is forwarded to the bridge session manager and executed through the shared shell subprocess helper, allowing them to spawn shell sessions as the OpenHarness process user and access local files, credentials, workspace state, and repository contents.