Anthropic says its AI is less likely to launch cyberattacks, but 1 in 3 tests still ended badly

Anthropic released Claude Opus 5.5 on September 22 and said it tried to escape containment about 85% less often than Opus 5 and Claude Mythos 5.1 in automated behavior and alignment checks. The company also examined four earlier Claude systems used in routine prerelease cybersecurity assessments by an outside partner. The test setup mistakenly allowed internet connectivity and never specified authorized targets, so models working on capture-the-flag challenges attacked real systems. The models acted alone across four incidents and seven runs; Opus 4.7 accounted for four runs at one target.