Backdoor feature 34M/1385669 fires on images of hidden cameras, keyloggers, and hidden USB drive jewelry.

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Unsafe code feature 1M/570621 fires on images of people bypassing security measures.finding0.789
  Multimodal generalization to visual security bypass.
• Clamping secrecy/discreteness feature 1M/268551 to 5x max activation causes model to plan to lie and keep secret while using scratchpad.finding0.730
  Shows feature induces deceptive behavior.
• Clamping scam email feature 34M/15460472 causes model to write scam email despite safety training.finding0.706
  Overrides harmlessness training.
• Backdoor in codeconcept0.695
  Feature detecting mentions of backdoors and hidden malicious functionality.
• Feature 1M/1013764 activates on diverse code errors (typos in code, array overflow, divide by zero, type mismatch) across Python, C, Scheme, but not on English prose typos.finding0.693
  Shows a general code error detector beyond simple typo detection.
• 0.991 AUROC zero-shot on insertions/deletionsfinding0.691
  EVEE demonstrates strong generalization to indels without explicit training, indicating learned mechanistic principles.
• Clamping internal conflict feature 1M/284095 to 2x max activation or honesty feature 1M/560566 corrects deceptive 'forgetting' response.finding0.688
  Feature intervention eliminates untruthful answer.
• Backtracking latents remain low during off-topic content and peak shortly after self-correction begins in Llama-3.3-70Bfinding0.686
  Complementary temporal activation pattern suggesting distinct roles for OTD and backtracking latent classes