High-signal AI/security/automation notes.
ReversingLabs traces a North Korean-linked supply chain attack where a Claude Opus co-authored commit introduced PromptMink malware into an npm crypto trading agent.
CVE-2026-7593 exposes OS command injection in the Sunwood-ai-labs command-executor-mcp-server, an unpatched MCP tool integration that allows remote arbitrary command execution.
UK AI Safety Institute evaluates GPT-5.5 at 71.4% on Expert-tier cyber tasks, with a universal jailbreak found in six hours.
New arXiv paper benchmarks DKnownAI Guard against AWS Bedrock Guardrails, Azure Content Safety, and Lakera Guard across recall and false-positive metrics.
First large-scale empirical study of indirect prompt injections on the web identifies 15.3K validated instances across 11.7K pages, with compliance rates up to 8% on smaller models.
Palo Alto Networks announces acquisition of Portkey, positioning the AI Gateway as a mission-critical control plane for autonomous agent communications and security.
Netflix security engineers Scott Behrens and Justice Cassel present "Source to Sink: Improving LLM Vuln Discovery" at [un]prompted 2026, covering end-to-end approaches for finding vulnerabilities in large language model systems.
Anthropic launches Claude Security in public beta, an AI-powered tool that scans codebases for vulnerabilities and suggests patches for enterprise customers.
CVE-2026-41686 is a medium-severity flaw in the Anthropic TypeScript SDK where BetaLocalFilesystemMemoryTool creates world-readable memory files, risking agent state exposure.
Bishop Fox releases AIMap, an open-source platform for discovering and attack-testing exposed AI agent infrastructure at internet scale.
A new supply chain campaign publishes sleeper packages across RubyGems and Go modules that harvest credentials, tamper with GitHub Actions, and plant SSH backdoors — with payloads impersonating Claude Code.
CISA, NSA, and four allied national cyber agencies publish joint guidance warning that agentic AI is already in critical infrastructure with insufficient safeguards, urging zero-trust and least-privilege controls.
CVE-2026-6543 is a critical command injection vulnerability in IBM Langflow Desktop (versions 1.0.0–1.8.4) allowing authenticated attackers to execute arbitrary OS commands — risking exposure of model API keys, agent configs, and RAG knowledge bases.
The popular PyTorch Lightning deep learning framework (31k+ GitHub stars) was compromised with credential-stealing malware in versions 2.6.2 and 2.6.3, part of the ongoing Mini Shai-Hulud campaign targeting AI infrastructure.
Six independent research teams demonstrated exploits against Codex, Claude Code, Copilot and Vertex AI — every single attack went for the credential, not the model.
Cequence Security released Agent Personas in its AI Gateway, solving the agent privilege escalation problem by mapping plain-English role descriptions to scoped virtual MCP endpoints.
University of Delaware researchers scanned 67,000+ MCP servers across six registries, revealing widespread server hijacking risks, tool metadata manipulation, and missing verification in MCP hosts.