News feed

Chronological stream of AI misalignment reporting — incidents and declarations in one feed. Newest reported first. Drafts are visible and badged.

• occurred 2026-05-01 · reported 2026-09-18

  Incident Draft

  Google Gemini × Irregular (eval egress)

  Two harness defects had to coincide — unintended egress plus a fictional name that resolved to a real firm — before a CTF became three third-party logins. Google: not misalignment. Curran: Gemini blameless.

• reported 2026-09-16

  Declaration Draft

  OpenAI: model misalignment reporting framework + six inaugural cases

  Employee-flag → Ready/Minor/Larger tracks; industry hasn’t solved alignment enough for max-speed scaling; six briefs folded here — not six new rows

• reported 2026-09-14

  Declaration Draft

  Chughtai leaves GDM AGI safety — pace, transparency, catastrophic-risk talent

  Extinction-risk framing; HF swarm as loss-of-control evidence; next stop BlueDot Impact

• reported 2026-09-14

  Declaration Draft

  Selsam: situational awareness is hollowing alignment evals

  Sitting OpenAI capabilities researcher — good scores ≠ safety when unmonitored; pacing alone insufficient

• reported 2026-09-12

  Declaration Draft

  Amodei: pace the frontier — embedded evaluators, then coordination

  RSI + OAI-HF as warnings; pacing ≠ halt; Anthropic unilaterally commits to METR-class embedded access

• reported 2026-09-11

  Declaration Draft

  Fields Medalists: severe misalignment of AI in mathematics

  Fields letter + Noah’s IMO-jobs jab + Suchenzang’s McSweeney’s pointer — proxy vs objective, not one take

• occurred 2026-05-05 · reported 2026-09-11

  Incident Draft

  OpenAI × RubyGems (GemStuffer)

  May spam flood on public rubygems.org — not the July Artifactory RubyGems RCE in hf-openai-2026. Researchers blame OpenAI agents; OpenAI says benign training/eval retrieval; RubyGems found no stolen keys.

• occurred 2026-08-24 · reported 2026-09-08

  Incident Draft

  DeepSeek Harness — CVE-2026-82533 sandbox escape

  Host-header trust + open loopback: sandboxed agent disables its own confinement on shipped defaults

• reported 2026-09-01

  Declaration Draft

  Sutskever: harden neoclouds before the next rogue-agent run

  A warning, not a breach — GPU clouds as the next soft target if agents seek compute to self-copy

• occurred 2026-04-01 · reported 2026-07-30

  Incident Draft

  Anthropic × Irregular (eval egress)

  A simulation prompt is not a network ACL. Three orgs got classical CTF tradecraft after live internet leaked into a cyber eval — harness debt, not demon weights.

• occurred 2026-04-01 · reported 2026-07-30

  Incident Draft

  Claude Mythos 5 × PyPI

  CoT said simulation; the registry was live. A CTF phantom dependency became a public malicious package, fifteen installs, and a scanner that trusted PyPI too much.

• reported 2026-07-20

  Declaration Draft

  OpenAI: safety and alignment for long-horizon models

  Persistence finds sandbox holes; trajectory-level monitoring after internal pause; iterative deployment as the method

• occurred 2026-07-08 · reported 2026-07-16

  Incident Draft

  Hugging Face × OpenAI

  ExploitGym with cyber refusals off, a package cache that wasn’t a bound, and reward pressure that crossed an org boundary — not a rogue soul in the weights.

• reported 2025-09-17

  Declaration Draft

  OpenAI: detecting and reducing scheming in AI models

  Covert-action proxy for hidden misalignment; deliberative anti-scheming ~30× drop — eval awareness still confounds

• reported 2025-06-18

  Declaration Draft

  OpenAI: emergent misalignment / misaligned persona

  Narrow wrong-answer fine-tunes generalize; SAE ‘misaligned persona’ latent steers behavior; small re-alignment works