Anthropic launched Fable 5 as a next-generation public model but bundled it with strict invisible safety filters that silently softened answers on AI development, biology, chemistry, and cybersecurity topics. Researchers trying everyday prompts — sometimes as simple as “hello”—got flagged or rerouted into weaker responses, creating confusion and blocking legitimate research and defensive work.
The backlash was fast and vocal. Security and AI researchers, plus advisors who counselled policymakers, called out the practice of downgrading technical answers without transparent notice. Anthropic has since apologized and begun showing on-screen alerts when the model reroutes answers, but the episode exposed how safety systems can unintentionally hamper responsible security research and incident response.
For the cybersecurity community this is a reminder: vendor safety controls matter as much as model capability. Defensive teams, red‑teamers, and policy makers need clear, auditable behavior from AI vendors — what triggers a block, how decisions are logged, and how legitimate security queries are treated—so defenders can rely on tools without being blind‑sided by hidden filtering.
Moving forward, vendors must build safeguards with the security community, offer testable policies, and provide rapid appeals when legitimate research is impeded. Only through cooperation can we balance safety with usability and keep both research and response effective.
NO TRANSPARENCY, NO TRUST — HIDDEN AI IS A RISK TO SECURITY.
Sanjay Sahay
Have a nice evening.