Kelsey Piper inside Anthropic: Claude alone now outperforms human+AI at plotting bioweapons
tszzl · x · 2026-10-22
- Kelsey Piper's piece in The Argument recounts a briefing at Anthropic HQ on AI-enabled bioweapons risk.
- Anthropic's red-team method: testers with only undergraduate chemistry knowledge tried using Claude to plan a mass-casualty event, scored by proximity to success.
- Striking finding: in early 2026, Claude working alone outperformed a human aided by Claude at drafting a workable plan — prompting an attendee to ask whether "the terrorist" is even needed anymore.
- Claude's capability has kept improving dramatically since, with charts shown at the briefing.
- Piper's verdict: the leading AI lab "has hopes, not plans," and she argues for regulation.
- Quoting the piece, tszzl counters that millenarianism drives people crazy, arguing convergent "end of the world" beliefs actually keep people sane and purposeful.
More from AGI Musings
- Defenders' AI, not attackers, found the vulns — fueling AI offense-defense debate — cephaloform · 2026-10-22
- 'Agent + cybersecurity will be HUGE': devs debate returning to security careers — max_paperclips · 2026-10-22
- Demis Hassabis: today's AI will look like nothing compared to what's coming in 1–2 years — VraserX · 2026-10-22
- AI Security Enters a Machine-Speed Arms Race, Human Patch Cycles Are Dead — srimisra · 2026-10-22
- Box CEO: Cybersecurity Will Be Where AI Creates the Most New Jobs — rakyll · 2026-10-22
- Digital minds don't live and die with each instance: a case for pattern-based identity — RileyRalmuto · 2026-10-22