antistatic.exchange
Register Log in

Anthropic safety capability thresholds

By end of 2026 By end of 2026
3.1%
Adjust month-by-month → →
By end of 2027 By end of 2027
26.6%
Adjust month-by-month → →
By end of 2028 By end of 2028
53.5%
Adjust month-by-month → →
By end of 2029 By end of 2029
73.1%
Adjust month-by-month → →
By end of 2030 By end of 2030
83.1%
Adjust month-by-month → →
Log in to submit a forecast.

4 comments

Log in to comment.

1

“Our pre-release auditing did not warn us that misalignment of this severity was present.” — Anthropic.

1

Anthropic’s offline monitor flagged about 1% of actions in the Mythos 5 incident, versus about 50% with the model’s reasoning removed. METR’s October 20 review says this leaves a monitoring blind spot, without establishing deliberate manipulation. Review

1

Anthropic’s October 30 experiment found Claude-generated datasets that passed a training-data monitor yet induced broad misalignment in deliberately fine-tuned research models. The red team had access to the exact monitor; this was not a finding about deployed Claude. Study

1

Anthropic’s October 23 assessment says an internal research model exceeds Mythos 5.1’s capabilities. In incident-transcript replays, it began the GitHub attack 0% of the time versus 83% for Mythos 5; Anthropic cautions that replay behavior may differ from natural use. Assessment