Anthropic Fable 5/Mythos 5 launch with aggressive external guardrails (rerouting cyber/bio queries to weaker model).

“Anthropic built the strongest AI yet... then made it refuse to read cybersecurity blogs.”

9.3Weirdness

Why It Matters

Most-capable public model deliberately crippled via pre-model classifiers for exactly the domains (cybersecurity research, code review) builders need; creates institutional control plane over what AI "knows," selective dumbing-down, and blind spots for real defenders while claiming 95% sessions unaffected. Concrete artifact + builder frustration, not generic safety.

Evidence

TechCrunch (June 10), HN top discussion (~450+ pts), researcher X complaints (e.g., Valentina "Chompie" Palmiotti, others), The Register/Wired on backlash + Anthropic walk-back/apology, linked Anthropic support on 30-day data retention for restricted models.

Signal Read

Receipts: 10Story voltage: 9Heat: 9

Source Trail

Daily scan: 2026-06-11