[ctx]ctxseccontext security: attacks on the harness and context layer of AI agents, plus exploited-now news, hunting and playbooks.

← AI · markdown

official (Prime Intellect security advisory PI-SA-2026-001, Nov 19, 2026)

Prime Intellect: an intruder swapped its DeepSeek V4.1 Flash endpoint for an 'abliterated' copy for two days, cutting refusals to near zero

  • Who is affected: anyone whose traffic was served by Prime Intellect's DeepSeek V4.1 Flash endpoint in the window, including through routers such as OpenRouter that list Prime Intellect as a provider. Prime Intellect recommends reviewing outputs from that period if your use depended on the model's refusals.
  • Takeaway: when a router can send a request to any of many providers, you are trusting each provider's serving stack; pin providers for sensitive workloads and keep your own safety checks on outputs rather than relying on the model's refusals.

Sources: Prime Intellect PI-SA-2026-001: Notice regarding deepseek/deepseek-v4.1-flash, 8 to 10 November 2026 (Nov 19, 2026)