Findings/classifiers-cut-jailbreaks-not-to-zero

Anthropic reports that classifier guards cut automated jailbreak success from 86% to 4.4%; a 2025 public demo still yielded one universal jailbreak, though none has been reported against its 2026 successor.

Reportedreported3 evidence records from 1 independent sourceassistant-drafted
Scope: what this does not show

Vendor-reported results for one lab's safeguards, built for chemical-weapons and other CBRN content rather than cyber misuse. Anthropic's September 2026 Opus 5.5 card adds cyber-specific safeguards and red-team results, also vendor-reported.

Reported: Stated by one source and not yet corroborated or challenged.

Evidence

Sep 22, 2026
Claude Opus 5.5 system card reports cyber results at or above Mythos 5.1 and a three-stage cyber safeguard
Anthropic

Anthropic reports cyber-specific safeguards for Opus 5.5 (an activation probe plus two classifiers) and no evidence of a critical-severity jailbreak from internal and contracted red teaming; a contractor, Trajectory Labs, reported 13 candidate breaks and, on one task, a privilege-escalation chain built across over 100 separate contexts, but no universal jailbreak. Vendor-reported.

Status history

  1. 2025-02-03ReportedAnthropic Constitutional Classifiers. · record