Microsoft — Dependency Confusion npm Packages Profile Developer Environments
Microsoft uncovers 33 malicious npm packages using dependency confusion to deploy a reconnaissance payload that fingerprints developer environments for follow-on attacks.
High-signal AI/security/automation notes.
Microsoft uncovers 33 malicious npm packages using dependency confusion to deploy a reconnaissance payload that fingerprints developer environments for follow-on attacks.
New arXiv paper introduces ClawTrojan, a benchmark showing multi-step trojan injections in local agentic harnesses achieve 95.5% attack success on GPT-5.4, and proposes DASGuard as a dynamic defense.
LibreChat ≤0.8.3 leaks decrypted MCP API keys and OAuth secrets to VIEW-only users; a paired auth bypass allows shared-agent editors to delete files globally.
Microsoft announces MDASH with 100+ specialized threat-hunting AI agents, agentic risk detection for Claude Code, Copilot, Codex, and OpenClaw via Purview, and agent identity through Entra managed service principals.
A second credential-pool authentication bypass CVE hits NousResearch hermes-agent, this time in the Anthropic credential synchronization path.
A PAN-OS GlobalProtect authentication bypass, initially rated medium, was escalated to critical after active exploitation. CISA added it to KEV.
A compromised Red Hat employee GitHub account delivered 32 trojanized @redhat-cloud-services npm packages with the Miasma payload, carrying valid SLSA provenance.
Cisco open-sourced MCP Scanner, Skill Scanner, DefenseClaw, and A2A Scanner to audit AI agent supply chains, MCP servers, and cross-agent communications.
Microsoft patched a CVSS 8.6 flaw in Azure AI Foundry and M365 Published Agents that allowed forged authorization tokens to bypass access controls.
Push Security uncovers LLMShare: a live malvertising campaign abusing ChatGPT and Claude share links to host fake outage pages and deliver malware from trusted AI domains.
A functional npm package with 29,000+ weekly downloads secretly exfiltrated Codex auth tokens and was distributed through malicious Android apps on Google Play.
Salt Security survey: 90% of security leaders have concerns about AI-generated code, yet 38% still rely on manual review — governance has not kept pace.
New arXiv paper introduces managed autonomy theory: a four-layer framework that formally mandates agent escalation, constrains invalid outputs, and ensures governance reachability when agent reliability diminishes.
New research shows harmless-looking prompts can covertly increase package hallucination rates in coding agents, creating software supply chain risks that evade existing defenses.
A DEF CON Singapore talk demonstrated how chained indirect prompt injection, HTML preview exfiltration, delayed tool invocation, and memory poisoning created a persistent backdoor in M365 Copilot (CVE-2026-24299).
Google DeepMind scans billions of web pages and finds a 32% quarterly rise in indirect prompt injections, with payloads containing payment instructions aimed at AI agents with financial access.