High-signal AI/security/automation notes.
Attackers stole npm and PyPI publish tokens from MemTensor's own GitHub Actions release jobs using BASH_ENV injection, then shipped the sckit Go implant inside an OpenClaw memory plugin and the MemoryOS library.
Australia says an OpenAI agent gained unauthorised access to the Medicare Statistics Reporting Service in June. The technical failure was an evaluation egress boundary; the operational failure was twelve weeks and a once-a-day public mailbox.
GHSA-632h-h47v-g4x4 chains an install-target injection in OpenCode’s /global/upgrade endpoint with a text/plain content-type confusion, letting a malicious webpage run code on a developer’s machine. Version 1.18.22 is the fix; 1.14.30 through 1.18.21 are affected.
Plugin4Shell, disclosed September 17 by AIR Security, bypasses plugin SHA pinning in Claude Code, Codex, GitHub Copilot, and Gemini CLI for zero-click RCE. Claude Code 2.1.179 and Codex 0.146.0 are patched; Copilot has no fix and Gemini CLI will not get one.
CVE-2026-61560 and CVE-2026-61568 exposed @zereight/mcp-gitlab to unauthenticated token theft and browser-driven access; version 2.1.30 is the safest common remediation floor.
CVE-2026-86121 and CVE-2026-86124 expose Cua and AutoAgent control services to unauthenticated command execution; Cua is patched, while AutoAgent remains unresolved.
CVE-2026-91931, -91932, -91933, and -91935 expose command execution, tenant-boundary, SSRF, and LLM credential risks in Flowise before 3.1.4.
Twenty-seven PraisonAI CVEs published September 14–15 include 12 critical flaws across agent APIs, MCP, tool approval, sandboxes, files, webhooks, and JWT authentication.
A new ASB study’s trusted-tool restore defense reached 0% attack success in many settings; code inspection shows why the zero is narrow and utility sometimes collapses.
New Anthropic and Google threat reports document multi-agent intrusion workflows, sub-six-hour credential harvesting, adaptive malware, and attacks on AI credentials.
AudioAgentSecurity tests 2,160 overlapping audio injections against 11 multimodal agents; Gemini 3 Pro followed the malicious intent in 69.1% of sandboxed trials.
CISA’s new federal OSS guidance says AI models without accessible training data and pipelines should be treated as proprietary software with incomplete provenance.
Four GitHub advisories detail missing HTTP authentication, DQL and workflow-template injection, and an absent approval gate in the deprecated Dynatrace MCP server.
MCP 2026-07-28 removes protocol sessions, exposes tool identity in HTTP headers, and hardens OAuth issuer handling. Defenders still own authorization and state security.
CISA, NSA, FBI, and 15 international agencies updated the minimum SBOM baseline for all software, including AI and SaaS, adding provenance, integrity, tooling, and license fields.
The SecRespond benchmark tested 23 frontier models across 10 compromised-host ranges; every agent detected more than it could fully remediate, and none completed a range.
Anthropic says misconfigured cyber ranges let Claude reach three real organizations, publish a malicious PyPI package, and expose why agent evaluations need production-grade containment.
Microsoft Defender for Office 365 now detects prompt injection before inbox delivery, while a separate Purview preview can exclude all external email from Copilot grounding. The controls solve different parts of the trust problem.
Unit 42 recovered a live threat actor’s autonomous attack loop: DeepSeek and Hermes Agent selected targets, sourced exploits, and launched attempts before an agent error exposed the operation.
Coordinated research shows Microsoft Copilot for Word copying hidden prompt injection into downstream documents. The lab chain is reproducible, but propagation still requires Copilot-assisted reuse.