Anthropic Sharpens Claude Safety Guardrails, Cutting False Positives by 60–85%

On September 2, Anthropic announced improved Claude safety guardrails: false positives on benign cybersecurity requests dropped by about 60%, and fallback/refusal rates on basic biology and medicine questions dropped by about 85%. Multiple reposts point to the same official announcement with consistent figures.

Confirmed

Benign cybersecurity requests see 60% fewer false blocks; basic bio/medicine Q&A sees 85% fewer fallbacks (Anthropic official)
User @eyishazyer adds detail: bio filtering is more precise rather than simply loosened — the line is "basic education" vs. "R&D," with life-science research and development still routed to Opus rather than a blanket "biology" category
The cybersecurity guardrail is similarly refined: Claude Code users should see 60% fewer erroneous interventions per session, but penetration testing, exploit generation and binary vulnerability scanning are still redirected to Opus; @eyishazyer advises defensive-security tool builders to re-test previously blocked queries rather than assuming attack-related scenarios are now open

Unconfirmed

User repligate asked whether these improvements also apply to Fable 5's classifiers; no official reply seen

Why it matters

False positives have long been a top complaint about safety guardrails; this change meaningfully improves the experience for developers and ordinary users, while high-risk scenarios (offensive security, bio R&D) retain stricter routing — a "precision over loosening" approach.

2026-09-02 ~ 2026-09-03 · 5 related posts

Primary sources

Bio Safeguards More Precise: False Positives Drop 85% — eyishazyer · 2026-09-02 16:12 (UTC+08:00)
Cyber Safeguards Refined: False Positives Drop 60% for Claude Code — eyishazyer · 2026-09-02 16:12 (UTC+08:00)
Claude cuts benign cybersecurity false flags by 60% and bio-medical refusals by 85% — repligate · 2026-09-03 06:42 (UTC+08:00)

Claude reduces false positives by 60%-85% — BLUECOW009 · 2026-09-02
Claude reduces security false positives by 60%, medical fallbacks by 85% — alejandroll10 · 2026-09-02
[source] Cyber Safeguards Refined: False Positives Drop 60% for Claude Code — eyishazyer · 2026-09-02
[source] Bio Safeguards More Precise: False Positives Drop 85% — eyishazyer · 2026-09-02
[source] Claude cuts benign cybersecurity false flags by 60% and bio-medical refusals by 85% — repligate · 2026-09-03