feed7.dev
Sign InStart Agent Brain
AnthropicEngineering PostOfficial Source

Announcements

Anthropic details Fable 5's cyber classifiers — pen testing and exploit dev are blocked even for legitimate use, with a wider false-positive margin — and drafts a five-level jailbreak severity scale (CJS).

Anthropic
Open Source Open MarkdownOpen JSON
Source Summary

**The gist** Anthropic published details on **Fable 5**'s cyber safeguards (**July 2, 2026**): classifiers sort requests into four buckets — prohibited (ransomware, C2, malware dev), **high-risk dual use** (pen testing, exploit development — blocked pending better access controls), low-risk dual use, and benign defensive work. It also drafts a **CJS jailbreak severity scale**, five levels scored on capability gain, breadth, ease of weaponization, and discoverability, built with **Amazon, Microsoft, and Google**.

Practical Implication

**Why it matters** If you do security work with Claude, expect refusals beyond the obvious: Anthropic deliberately **widened the safety margin**, so some benign and low-risk requests get blocked as accepted false positives. Defensive tasks — secure coding, debugging, incident response, **malware reverse engineering** — are meant to stay open.

Agent-Ready Context
**The gist** Anthropic published details on **Fable 5**'s cyber safeguards (**July 2, 2026**): classifiers sort requests into four buckets — prohibited (ransomware, C2, malware dev), **high-risk dual use** (pen testing, exploit development — blocked pending better access controls), low-risk dual use, and benign defensive work. It also drafts a **CJS jailbreak severity scale**, five levels scored on capability gain, breadth, ease of weaponization, and discoverability, built with **Amazon, Microsoft, and Google**.

**Why it matters** If you do security work with Claude, expect refusals beyond the obvious: Anthropic deliberately **widened the safety margin**, so some benign and low-risk requests get blocked as accepted false positives. Defensive tasks — secure coding, debugging, incident response, **malware reverse engineering** — are meant to stay open.

**Watch out** This is an **early draft**: classifiers will shift with feedback, and CJS scores can be raised but **never lowered**. Legitimate pentesters have no unlock path yet — better access controls are promised, not shipped.
Context Map
modelsecurity#model-selection
Uncertainty
This is an **early draft**: classifiers will shift with feedback, and CJS scores can be raised but **never lowered**. Legitimate pentesters have no unlock path yet — better access controls are promised, not shipped.