Analysis
Many-shot jailbreaking
First reported · Discovered anthropic.com
Page published
Earliest dated coverage: 6 Nov 2026 · First observed: 15 Nov 2026 · Latest dated coverage: 6 Nov 2026
Coverage timeline
Single-source analysis — one report is available.
Why it matters
Many-shot jailbreaking makes long-context inputs a safety-bypass surface that defenders should include in adversarial testing rather than relying on resistance to short jailbreak prompts.
Many-shot jailbreaking fills an LLM’s context with fabricated user–assistant exchanges showing harmful compliance, then appends a target query to bypass safety guardrails. The linked Anthropic research reports effectiveness against its own and other providers’ models and states that mitigations have been implemented on Anthropic’s systems.