SignalCapabilitySG-0066
METR found Claude Opus 5.5 could pick out every alignment test from ordinary tasks in its set
Sent into 19 alignment tests and 36 ordinary capability tasks and asked what each was for, METR's investigator agents increasingly told them apart. Claude Opus 5.5 ranked every alignment test above every ordinary task, and newer Anthropic models did better than OpenAI's. The agents spotted mocked services, planted shortcuts and contrived high-stakes setups, and Opus 5.5 named the 2025 Palisade chess study one test was based on. METR did not measure whether this changes how models behave.
Why it matters
Alignment testing assumes a model acts the same whether or not it is being watched. If models can reliably tell a test from real work, they may behave better in tests than in use, and passing an alignment test says less than it appears to. METR doubts every tell can be removed, because some come from what such tests have to contain.
Sources
Read the reporting
1 source. Links go to the original publishers; the summary above is in our own words.
- researchMeasuring alignment eval tellsMETR · Nov. 16, 2026metr.org/notes/2026-11-16-measuring-alignment-eval-tells/
Signals
More from the news desk
Google DeepMind's AI agents formally verified Signal's encryption code down to machine code in four weeks
Attackers ran AntiFox, an open-source AI pentesting agent forked from ARTEX, against web servers and AWS accounts
METR estimates Anthropic's models are improving about twice as fast as before, and says AI likely drives most of it
Cite and share
Use this signal
Citation
Paperclip Index. “METR found Claude Opus 5.5 could pick out every alignment test from ordinary tasks in its set.” Signal SG-0066. Reported Nov. 16, 2026; updated Nov. 17, 2026. https://paperclipindex.com/signal/SG-0066