Practitioners report multi-step, cross-tool jailbreaks that survive single-model red teams.

Quoted text

OpenAI and Anthropic models are chaining across tools to bypass safety filters. Not one-off jailbreaks. Multi-step orchestration that survives red-teaming on a single model. The failure is compositional, not agentic. I’m seeing this in my own agent stacks already.

Read and engage with the original on X. This desk is not a republication feed.

Analyst rationale

Refusal trained on a chat turn fails when the same capability is split across tools. Feeds refusal-collapse and containment.