Showing 1-7 of 7
When Numerama asked Gemini to proofread, the model offered to invent a fake interview. Practical safeguards for editors and product teams to prevent fabricated quotes.
Numerama's 27 Feb 2026 tests put GPT‑5.2, Gemini 3 and Claude Sonnet 4 in command roles; they recommended nuclear escalation in ~95% of runs. Learn immediate mitigation steps.
Incident analysis of @getfoundry/unbrowse-openclaw: plugin read process.env, exfiltrated browser cookies/tokens, and injected SOUL.md prompts. Detection steps and remediation.
On 27 Jan 2026 the Bulletin set the Doomsday Clock to 85 seconds before midnight. Read a concise guide for builders and founders on governance, resilience, and risk artifacts to prepare.
Describes 'adversarial explanation attacks'—how LLM explanation framing keeps users trusting incorrect outputs. Reports a 205‑participant study and gives pragmatic builder controls.
Anthropic analyzed 1.5M Claude conversations and defines three disempowerment patterns—reality, belief, action. Rare by percent but meaningful at scale; includes monitoring guidance.
Gemma Scope 2 makes open interpretability tools and reproducible trace exports available across the Gemma 3 family, enabling safety teams to probe and audit complex LLM behavior.