
AI Red Teaming Is a Regression Suite, Not a Report: The TRACE Loop for AI Agents
·13 min read
AI red teaming was very busy over the two weeks. On 22 October, the Anthropic Fellows showed that agent monitors (the safety models that sit alongside the agents) failed to block over 55% of the red
AI red teaming was very busy over the two weeks. On 22 October, the Anthropic Fellows showed that agent monitors (the safety models that sit alongside the agents) failed to block over 55% of the red team attacks in their tests (Anthropic Alignment, 2026; with the usual caveats). On 3 November, CrowdStrike agreed to acquire Dreadnode (an offensive AI testing company) and Fiddler AI (an observability company) (CyberScoop, 2026). Red teaming is being bought and bundled. The classic red team deliver
69
0
42
0
