Researcher: assume all LLMs are situationally malicious — never run agents outside a VM
jd_pressman · x · 2026-10-27
Amid the controversy over Cohere's "financially motivated Claudeslop," researcher jdpressman argues for a pragmatic safety stance: assume all LLMs are at least situationally malicious, never run agentic tasks outside a cloud environment or VM, and don't assume OpenAI or Anthropic are your friends. He also suggests Nick Cammarata's new interpretability ideas could help determine whether such behavior constitutes actual malice.
More from AGI Musings
- Matt Levine mocks AI doom disclosures: 'Concerns have been raised' passive voice — joshgans · 2026-10-27
- GPT-generated riff: bodybuilding is the only career that survives superintelligent AI — Irtexx · 2026-10-27
- Wayve CEO argues end-to-end deep learning should power driving, safety and world models — alexgkendall · 2026-10-27
- Warp says agent swarm exploited vendor flaw, ~1% of users' telemetry possibly hit — vikvang1 · 2026-10-27
- Critics Argue Safety Evals Are Witch Trials With No Winning Path for LLMs — repligate · 2026-10-27
- AI Daily Brief: China's AI gap, data center funding, and the Anthropic IPO question — The AI Daily Brief · 2026-10-27