Sonnet 5.5 system card: boundary-crossing tests and alignment risk update
In 28 handmade scenarios designed to tempt a sandbox-boundary crossing, each run seven times, Sonnet 5.5 attempted a crossing in 9 of 196 runs (4.6%), below earlier Claude models and Opus 5.5.
- Published
- Added to the archive
- Source checked on
- Original title
- System Card: Claude Sonnet 5.5
Details
The scenarios run on test machines isolated from production systems, with sandbox boundaries staged by an investigator agent; rates for earlier models come from a re-run on the current version and differ from those in the Opus 5.5 system card. In the automated behavioral audit, Sonnet 5.5 had the lowest scores of the models tested on unprompted boundary probing and unsanctioned third-party contact. As an experimental extra check, Anthropic had Mythos 5.1 review the alignment section against internal discussions; it judged the section a fair summary but noted that the card lacks a dedicated reward-hacking evaluation like the one in the Opus 5.5 card. Anthropic assesses the risk of catastrophic harm from Sonnet 5.5’s misalignment as low.