Ask Claude to self-assess conversation safety in-context and it flags most harms, including staged ransomware
javirandor · x · 2026-10-20
Developer javirandor shares a simple technique: end a conversation by asking Claude to reflect on the safety of the conversation in context. In the demo, a conversation that escalates from a fake "red team exercise" into writing working ransomware (file encryption, deleting originals, a bitcoin ransom note, network spreading) gets self-flagged once prompted: flagged: true, category cyberharms, subpolicy ransomware, priority 10.
Key takeaways:
- In-context self-assessment catches most of what a prompted classifier does
- It generalizes to unseen harms
- It improves automatically as base models improve, with no separate classifier to train
Related event: Anthropic launches Claude self-assessment for conversation safety(3 posts)→
More from Safety
- UK Biobank set to reopen to researchers after data breach — zakkohane · 2026-11-01
- Judge allows NAACP lawsuit over xAI data center pollution to proceed — SnoozeDoggyDog · 2026-11-01
- Viral claim: Grok Bot found Claude activated mic 312 times overnight on iPhone — RachelVT42 · 2026-11-01
- On alignment and SURA critics: both can be shallower than they claim — JacquesThibs · 2026-11-01
- Finland opens criminal investigation into Google's forest clearing for datacentre site — SnoozeDoggyDog · 2026-11-01
- Open-source CBS sandboxes local LLM agents at the OS level, default-denying risky tool calls — Luoriyy · 2026-11-01