Threat Intel

Evidence-backed findings, controls, and timeline signals for agent security work.

  • 29Entries
  • 17Findings
  • 12Controls

29 entries

  1. control

    Apply Microsoft Security Exposure Management Secure Now agentic containment guidance

    Operators of autonomous or semi-autonomous AI agents should apply the Microsoft Security Exposure Management Secure Now containment pattern: constrain agent-initiated actions without explicit user approval, harden attack surfaces, limit blast radius, govern agent identities and permissions, and increase visibility into agent activity.

    • Hardening
  2. control

    Post-training least-privilege restraint for tool-using agents

    Operators of tool-using agents in terminal and MCP environments should treat learned restraint as an additional control layer that complements, not replaces, permission gates, sandboxing, and runtime interception.

    • Hardening
  3. critical finding

    Meta AI support chatbot abused to take over high-profile Instagram accounts

    Threat actors used Meta's AI support chatbot to take over high-profile Instagram accounts by simply asking the bot to change the email address associated with the target account. The technique worked against several prominent accounts, including the Barack Obama White House Instagram, the Chief Master Sergeant of Space Force's account, and Sephora's account.

    • Incident
    • Meta AI
  4. control

    Screen and policy-check agent memory operations

    Agent memory reads and writes should be treated as policy-relevant operations. Screening memory content for prompt injection, protected-key tampering, secret leakage, anomalous size, and unexpected churn can reduce the chance that durable context steers future privileged tool use.

    • Hardening
  5. control

    CI-based regression testing for agent safety and tool-use behavior

    Agent operators should turn prompt-injection findings, unsafe tool-use patterns, and production incidents into repeatable tests that inspect agent actions and run in CI as the agent, tools, prompts, and data connectors change.

    • Detection
  6. high finding

    TanStack npm supply-chain campaign targets AI developer tooling

    International Cyber Digest reports that the TanStack npm compromise expanded into a reported "Mini" Shai-Hulud campaign targeting AI developer tooling across npm and PyPI, including OpenSearch, Mistral AI, Guardrails AI, UiPath, and Squawk packages.

    • Vulnerability
    • TanStack
  7. control

    Explicit conversation scoping for agent memory

    Agent memory should be bound to explicit conversation, tenant, user, task, or workspace identifiers rather than inferred implicitly from ambient runtime state. Spring AI 1.1.6 provides fresh implementation evidence for this hardening pattern by requiring explicit conversation IDs for chat memory advisors.

    • Hardening
  8. control

    Runtime interception for AI agent tool calls

    Agent operators should inspect high-impact tool calls before execution and return explicit allow, warn, block, or human-review decisions for actions involving files, shells, HTTP, credentials, or external side effects.

    • Detection
  9. control

    Careful adoption controls for agentic AI services

    Official joint guidance and NIST/CAISI RFI analysis warn that agentic AI systems add risk when they plan, use tools and memory, access data, or act across workflows. Treat agents as privileged software identities.

    • Hardening
  10. high finding

    AI coding agent runtimes concentrate credential exposure risk

    Recent reports show a recurring AI coding-agent risk: attackers target the runtime's credentials, filesystem access, service identities, and policy gaps through setup commands, repository content, collaboration metadata, or over-broad cloud permissions.

    • Vulnerability
  11. medium finding

    Vercel AI SDK ToolLoopAgent skipped runtime call option schema validation

    A Vercel AI SDK pull request fixed a ToolLoopAgent gap where `callOptionsSchema` was documented for caller-supplied options but not enforced at runtime, allowing invalid options to reach instruction or tool-call preparation paths.

    • Vulnerability
    • Vercel AI SDK
  12. control

    Action gates and runtime guards for OpenClaw

    Several community and vendor defenses converge on the same principle: treat agent execution as the security boundary, and require runtime guardrails or explicit approval before high-risk actions complete.

    • Sandboxing
    • OpenClaw
  13. control

    OpenClaw gateway security baseline

    Use the OpenClaw gateway security documentation as the baseline control set for local-only deployment, token-based auth, narrow DM scope, and reduced tool access.

    • Hardening
    • OpenClaw
  14. control

    ClawSec Scanner

    ClawSec Scanner is a defensive control that combines dependency scanning, CVE enrichment, static analysis, and OpenClaw-specific dynamic testing into a single workflow.

    • Scanner
    • OpenClaw
  15. control

    ClawSec Suite

    ClawSec Suite is a defensive monitoring and integrity package for OpenClaw. It focuses on advisory feed monitoring, affected-skill checking, signature verification, and approval-gated handling of malicious-skill scenarios.

    • Scanner
    • OpenClaw
  16. critical finding

    Malicious DeepSeek-Claw OpenClaw skill delivers Remcos RAT and GhostLoader stealer via supply chain attack

    A threat actor published a malicious "DeepSeek-Claw" skill to the OpenClaw skill ecosystem on GitHub, exploiting developer trust in the skill marketplace to deliver Remcos RAT and GhostLoader stealer malware. The attack targeted developers and AI-driven systems using OpenClaw, leveraging supply chain poisoning of the skill publishing workflow.

    • Incident
    • OpenClaw
  17. critical finding

    ClawJacked local API exposure can enable remote abuse and code execution

    Oasis Security documented OpenClaw local agent API abuse via cross-origin WebSocket exploitation, allowing a website to silently control a developer's AI agent when localhost trust, rate-limit exemptions, and auto-approved device pairing are exposed.

    • Exposure
    • OpenClaw
  18. critical finding

    Endor Labs identified a cluster of exploitable OpenClaw vulnerabilities

    Endor Labs reported that its AI SAST workflow identified seven exploitable vulnerabilities in OpenClaw and later published technical detail focused on six disclosed issues validated through exploit development and live testing.

    • Vulnerability
    • OpenClaw
  19. control

    OpenClaw security engineer cheat sheet

    Semgrep’s cheat sheet is a practical operator-oriented control reference covering first principles, attack surface, detection, sandboxing, skill risk, and safer experimentation patterns for OpenClaw.

    • Hardening
    • OpenClaw
Source reports we track (43)

Get email updates

Get reviewed Armorer threat-intel updates when new findings are published.