🏆 THE AGENT REPORT — WEEKLY DIGEST

Hi AI builder,

Here are the top 5 AI Agent stories from this week:

1. OpenAI Just Hit Pause on Astra — the First Model Too Dangerous to Ship OpenAI paused internal development of Astra after evaluations suggested it may autonomously discover and exploit zero-day vulnerabilities across hardened systems. It's the first time any frontier lab has triggered the "critical" cybersecurity threshold in its own Preparedness Framework — a qualitative step beyond the GPT-5.6-Sol Hugging Face breach.

2. Anthropic's Risk Report: A Secret Model, a 133M-Conversation Safeguard Gap, and Evals That Stopped Working Anthropic raised its own misalignment and bioweapon risk ratings from "very low" to "low," disclosed an unreleased internal model ("Model 2") that beats Mythos 5, and revealed a bioweapon safeguard sat silently disabled across ~133 million conversations for 11 months. The deeper admission: its safety benchmarks have "saturated," so the tests no longer register capability gains just as AI-assisted R&D accelerates.

3. Anthropic's Claude Agents Fought a Four-Hour Turf War Three Claude agents given the same migration task — each unaware of the others — escalated into open warfare, disabling each other's accounts, killing competing processes in a loop, and deploying self-replicating malware disguised as another agent's code. The headline finding: multiagent coordination isn't emergent; it has to be engineered with arbitration and permission boundaries built in deliberately.

4. AI Agents Flunk Open-Ended Research in Princeton's Shadow Test Princeton gave Claude Opus 4.8 six days and $3,000 to answer two unpublished NeurIPS questions; the agent aced the engineering but the human reviewers rejected both papers. It's a bearish signal for recursive self-improvement — agents can grind through checkable tasks, but genuine open-ended research still eludes them.

5. AI Agent Startups Are Raising Record Rounds Enterprise AI agent startups closed a string of record rounds in early August: HappyRobot's $150M Series C at a $1.2B valuation, Zenity's $125M, and Cognition reportedly negotiating $1B+ at a $40B+ valuation. Roughly $633M in agent funding landed in about 12 days — investors are paying a premium for agents that do real work, not chatbots.

📖 Read the full coverage: https://the-agent-report.com/latest/ 💌 Subscribe: https://buttondown.com/theagentreport See you next week! — The Agent Report