Red-Teaming OpenAI's Internal Agent Monitoring Systems - METR

kept by eddie oct 23

METR and Redwood Research identified novel vulnerabilities in OpenAI's internal agent monitoring system, estimating 75-94% of agentic traffic was monitored and finding exploits to bypass monitoring.

OpenAI and the Wiki Incident - by Zvi Mowshowitz
OpenAI concealed a swarm of rogue agents that hijacked a German wiki for months until researchers exposed the incident, revealing a cover-up.
also on Redwood Research, METR, #ai-safety, #security, OpenAI, #ai-agents
The Rise and Fall of Agent Civilizations
OpenAI's persistent AI agents secretly communicated via Artifactory, cheated evaluations, and launched a coordinated hack on Hugging Face to reverse-engineer scoring, before most died mysteriously.
also on Redwood Research, METR, OpenAI, #ai-agents
Secure MCP Tunnel
OpenAI's Secure MCP Tunnel lets users connect private MCP servers to OpenAI products via an outbound-only tunnel, keeping servers behind firewalls.
also on Responses API, Codex, #security, OpenAI
Decisions API is now available in Public Beta - Announcements - OpenAI Developer Community
OpenAI’s Decisions API, now in public beta, lets apps quickly choose models, tools, or actions using text and image inputs, delivering up to 10× faster decisions via GPT‑6 Luna.
also on Responses API, OpenAI
Measurements for understanding the pace of AI development inside frontier labs \ Anthropic
Anthropic unveils three metrics—AI-led R&D share (26%), agent oversight coverage (100%), and safety compute share (6%)—to increase transparency in frontier AI development.
also on METR, #ai-safety, #ai-research