threatdeck
Live 14:45 UTC ·

All decks

ThreatDeck — 2026-11-20

12 cards today · 4 actively exploited or must-patch · 6 affecting AI tooling.

  1. 01

    Act Now Critical KEV CVE-2026-40217 AI-assisted

    CISA adds BerriAI LiteLLM CVE-2026-40217 remote code execution flaw to KEV catalog

    CISA added CVE-2026-40217, an unprotected alternate channel vulnerability in BerriAI LiteLLM, to its Known Exploited Vulnerabilities catalog on 2026-11-18. The flaw could let remote attackers execute arbitrary code via bytecode rewriting at the /guardrails/test_custom_code URI. The remediation due date is 2026-12-02. Use in ransomware campaigns is listed as unknown.

    Why it matters. LiteLLM is commonly deployed as an LLM gateway, so exposed instances could give attackers code execution on hosts that hold model API keys and routing configuration. Inclusion in the KEV catalog indicates the flaw is known to be exploited.

    Do this. Apply the mitigations or fixed version described in the BerriAI advisory GHSA-wxxx-gvqv-xp7p.

    2026-11-18 · score 77 · CISA Known Exploited Vulnerabilities

  2. 02

    Act Now Critical KEV CVE-2026-90817 AI-assisted

    CISA adds REDCap code injection flaw CVE-2026-90817 to KEV; unauthenticated RCE

    CISA added CVE-2026-90817, a code injection vulnerability in Vanderbilt University's REDCap, to its Known Exploited Vulnerabilities catalog on 2026-11-18. An unauthenticated attacker could execute arbitrary code on the server using crafted HTTP requests to an unintended controller route from a public survey context. Exploitation requires knowledge of a valid public survey hash. The federal due date is 2026-11-21.

    Why it matters. Organizations running internet-exposed REDCap instances, such as research and healthcare institutions, risk server compromise and exposure of the data held in REDCap. The flaw is listed as exploited, and the remediation deadline is only three days after it was added.

    Do this. Apply vendor mitigations or updates for REDCap as directed by the REDCap community resources.

    2026-11-18 · score 63 · CISA Known Exploited Vulnerabilities

  3. 03

    Act Now Critical KEV CVE-2026-90436 AI-assisted

    CISA KEV: Apache Axis unsafe reflection flaw (CVE-2026-90436) allows unauthenticated RCE

    CISA added CVE-2026-90436, an unsafe reflection vulnerability in Apache Axis, to its Known Exploited Vulnerabilities catalog on 2026-11-18. An unauthenticated remote attacker could execute arbitrary code with a crafted SOAP fault message, if a suitable class is on the application classpath. CISA notes the product could be end-of-life and sets a due date of 2026-11-21.

    Why it matters. Internet-facing Java applications that use Axis for SOAP handling could be taken over without credentials. Because Axis may be end-of-life and embedded as a library, affected systems may be hard to identify and may lack patches.

    Do this. Inventory applications and dependencies that bundle Apache Axis, including transitive library use, and identify which are internet-exposed.

    2026-11-18 · score 62 · CISA Known Exploited Vulnerabilities

  4. 04

    Act Now Critical KEV CVE-2024-10441 AI-assisted

    CISA adds Synology DSM CVE-2024-10441 remote code execution flaw to KEV catalog

    CISA added CVE-2024-10441 to its Known Exploited Vulnerabilities catalog. The flaw is an improper encoding or escaping of output issue in the system plugin daemon in Synology DiskStation Manager (DSM) and BeeStation OS (BSM). It could allow a remote attacker to execute arbitrary code via unspecified vectors. The due date is 2026-11-21.

    Why it matters. Organizations running Synology DSM or BeeStation devices exposed to the network could allow remote code execution by an attacker. Inclusion in the KEV catalog indicates exploitation, and federal agencies face a short remediation deadline.

    Do this. Apply the fixes described in Synology advisory Synology_SA_24_20 to all DSM and BeeStation OS devices.

    2026-11-18 · score 62 · CISA Known Exploited Vulnerabilities

  5. 05

    AI Stack Alert Critical CVE-2026-77359 AI-assisted

    CVE-2026-77359: Jupyter MCP Server before 1.0.3 lacks origin checks on /api/connect and /api/stop

    Jupyter MCP Server prior to 1.0.3 validates Host and Origin for /mcp in standalone streamable-http mode but not for the state-changing /api/connect and /api/stop routes. A malicious website using DNS rebinding can send a browser-shaped request that /api/connect accepts, replacing runtime_url and document_url values and starting a kernel client against attacker-controlled infrastructure. The issue is fixed in version 1.0.3.

    Why it matters. Organizations running Jupyter MCP Server in streamable-http mode, especially on loopback, LAN, VPN, or published-container listeners, can have notebook traffic redirected and kernels stopped by a visiting browser. The source reports a CVSS base score of 9.3.

    Do this. Upgrade Jupyter MCP Server to version 1.0.3 or later.

    2026-11-19 · score 43 · NVD (AI keyword search)

  6. 06

    AI Stack Alert High CVE-2026-77318 AI-assisted

    Jupyter MCP Server before 1.0.3: unauthenticated /api/connect and /api/stop routes bypass MCP_TOKEN

    Jupyter MCP Server prior to 1.0.3 enforces MCP_TOKEN on /mcp in standalone streamable-http mode but leaves the /api/connect and /api/stop management routes unauthenticated. An unauthenticated network caller can use /api/connect to replace runtime and document URLs and tokens, and can stop the current notebook kernel. The issue is fixed in version 1.0.3 and has a CVSS base score of 7.3.

    Why it matters. Any network-reachable deployment in this mode can be pointed at attacker-controlled Jupyter backends, causing notebook traffic to be sent there. Attackers can also stop active notebook kernels without authentication.

    Do this. Upgrade Jupyter MCP Server to version 1.0.3 or later.

    2026-11-19 · score 37 · NVD (AI keyword search)

  7. 07

    AI Stack Alert High CVE-2026-133161 AI-assisted

    CVE-2026-133161: Pydantic AI Harness ToolOutputLimits strip_ansi quadratic ReDoS stalls event loop

    With strip_ansi=True, Pydantic AI Harness's ToolOutputLimits capability removes ANSI escape sequences from tool output using a pattern whose time is quadratic in line length. A tool result of a few hundred kilobytes with many unterminated escape sequences can block the event loop for tens of seconds to minutes. The advisory gives a CVSS base score of 7.5.

    Why it matters. Applications that run tools whose output an attacker can influence, such as shell commands, file reads, fetched pages, or MCP servers, can have every request and agent run in the process delayed. The stripping runs before the capability's own size limits apply, so those limits do not bound the cost.

    Do this. Upgrade pydantic-ai-harness to 0.70.0 or later.

    2026-11-19 · score 37 · NVD (AI keyword search)

  8. 08

    AI Stack Alert High CVE-2026-130008 AI-assisted

    CVE-2026-130008: sqlite-mcp read-only mode bypass via execute_query and describe_table (fixed in 0.1.2)

    sqlite-mcp is a Model Context Protocol server for SQLite. Prior to version 0.1.2, its default read-only mode can be bypassed through execute_query and describe_table, allowing an ATTACH DATABASE statement that opens files read-write. The server has no built-in authentication and listens on all interfaces at port 8080 by default.

    Why it matters. Any client that can reach the server port can invoke these tools, so an attacker may modify a protected database, read other SQLite files, or create SQLite-format files writable by the server process.

    Do this. Upgrade sqlite-mcp to version 0.1.2 or later.

    2026-11-18 · score 37 · NVD (AI keyword search)

  9. 09

    AI Stack Alert High CVE-2026-130007 AI-assisted

    CVE-2026-130007: k8s-mcp-server before 1.1.0 exposes unauthenticated Kubernetes tool invocation over SSE

    k8s-mcp-server is a Model Context Protocol server for managing Kubernetes clusters. Prior to version 1.1.0, SERVER_MODE defaults to SSE on 0.0.0.0:8080 without an authentication option, and the POST /message handler dispatches tools/call without an authorization or Origin check. Read-only mode is disabled by default, so write-capable Kubernetes tools are available.

    Why it matters. Any client that can reach the listening port can enumerate and invoke Kubernetes tools without credentials, using the server's configured cluster privileges. This could allow unauthorized changes to cluster workloads.

    Do this. Upgrade k8s-mcp-server to version 1.1.0 or later.

    2026-11-18 · score 37 · NVD (AI keyword search)

  10. 10

    AI Stack Alert High Unconfirmed AI-assisted

    obsidian-sync-mcp below 0.5.7: missing Host/Origin validation enables DNS rebinding in no-auth mode

    The npm package obsidian-sync-mcp, versions below 0.5.7, does not validate Host or Origin headers. A public website can use DNS rebinding to drive a victim's Obsidian vault through the MCP server when no-auth mode is enabled. The advisory is rated high and is fixed in version 0.5.7.

    Why it matters. Any user running an affected server in no-auth mode may have their local Obsidian notes read or modified by a malicious web page the user visits. Attackers gain access to vault contents without credentials.

    Do this. Upgrade the npm package obsidian-sync-mcp to version 0.5.7 or later.

    2026-11-19 · score 36 · MCP Server Watch: advisories

  11. 11

    New Technique Medium TD-2026-0136 AI-assisted

    BlockJack: Fault attack redirects vLLM KV cache reads across tenants

    BlockJack is a fault attack on the paged KV cache address translation in LLM serving frameworks such as vLLM. By corrupting block-table entries, it redirects a request's KV cache reads to blocks assigned to another tenant's request, without modifying model parameters, prompts, or cached KV states. The authors tested seven models across Qwen, Llama, Gemma, and Phi families, and report that a faulted mapping exposes victim context.

    Why it matters. Multi-tenant LLM serving hosts that share a KV cache can leak one tenant's context to another and let an attacker steer another tenant's answers. The integrity attack caused up to 41.1% of responses on a 248-scene medical benchmark to recommend less urgent care.

    Do this. Apply the vLLM and Sonar patches for block-table assignment validation once available in your deployed version.

    2026-11-19 · score 23 · arXiv (AI security papers)

  12. 12

    New Technique Medium TD-2026-0134 AI-assisted

    ToolPulse: cache side-channel attack recovers tool-call workflows of LLM coding agents

    The paper presents ToolPulse, a cache-based side-channel framework that monitors tool invocations by agents running on the user's machine while inference stays remote. The authors report that Flush+Reload enables accurate, real-time tool monitoring, while Prime+Probe incurs prohibitive overhead. The Tool Recovery Attack reconstructed tool workflows across 11 LLMs on Codex and Claude Code, with the lowest F1 score above 93%.

    Why it matters. Organizations running tool-using LLM agents on shared hosts may leak which tools and workflows their users run, and which tasks are sensitive, to a local attacker with unprivileged cache access. The attack relies on shared system libraries, so the exposure is not limited to the model vendor.

    Do this. Treat local tool execution by LLM agents as an observable surface on multi-tenant or shared hosts.

    2026-11-19 · score 23 · arXiv (AI security papers)

Following

Follow the vendors, products, AI tools, techniques and actors you care about. Matching stories, CVEs and exploited bugs are marked, a Following tab appears in the feed, and returning visits show what changed. Saved only in this browser.

Presets

Get updates

A personal RSS feed of new stories on everything you follow. Paste it into any feed reader, Slack or Teams RSS app.