Anthropic gives Claude Sonnet 5.5 flagship cyber safeguards — and routes risky requests to the older Sonnet 5

#What happened

Anthropic's Sept 28 launch of Claude Sonnet 5.5 ships with cybersecurity safeguards the company previously reserved for its most capable models — the first Sonnet to get them. Anthropic says Sonnet 5.5's cyber skills are a "large improvement" over Sonnet 5 and comparable to Opus 5. Higher-risk requests (penetration testing, exploit generation) are detected and visibly handed to the older Sonnet 5; routine bug fixing is unaffected, though Anthropic warns of increased refusals on benign cybersecurity tasks.

System-card numbers: with safeguards off, Sonnet 5.5 averaged 11.53 capability flags and an 80% capture rate on ExploitBench, produced 178 full arbitrary-code-execution exploits (vs 1 for Sonnet 5), completed 46.1% of CyScenarioBench challenges (vs 0.7%), and 50 control-flow hijacks on the Binary Exploitation Benchmark (vs 3). Safeguards run in three stages (probe on internal activations, lightweight on-model classifier, separate trained LLM classifier) with 99.43% recall and a 21.0% attack success rate on the rewind-attacker robustness evaluation, improved from 57.2% for Sonnet 5. API developers must opt in to automatic fallback.

#Why it matters

The industry's dangerous cyber capability is moving down-market in weeks — this is the first time flagship-level guards ship on a mid-tier model. Anthropic is shifting from limiting what models can do to gating who can use the sharpest cyber skills: an expanded Cyber Verification Program gives vetted defenders tiered access to advanced capabilities on Sonnet 5.5, Opus 5.5, and Claude Mythos. Sonnet 5.5 is also the first Sonnet with anti-distillation classifiers and expanded "preserved thinking" so reasoning cannot be decoupled from the account. Watch Haiku 5.5 — if the cheapest tier also needs cyber gates, "frontier" will no longer mark where AI risk sits.

#Sources

1. TheStreet: Anthropic Claude Sonnet 5.5 Cybersecurity
2. VentureBeat: Anthropic Launches Claude Sonnet 5.5
3. Unite.AI: Anthropic Releases Claude Sonnet 5.5
4. FoneArena: Anthropic Claude Sonnet 5.5 Features

Verified September 30, 2026.