The Intake — Sunday, August 9, 2026

If you build with biology or health content in Fable 5 and have been hitting refusals, the classifier was rewritten on August 7. Anthropic says the revision cut blocked queries by approximately 85%.

Anthropic rewrote the classifier's "constitution" with expert guidance to better distinguish everyday health and educational questions from dual-use research. Queries touching virology, toxicology, molecular design, and drug design continue to route to Opus 5. The same day, Stanford University and the Arc Institute published a paper in Science. The paper showed their Evo 1 and Evo 2 models can design fully functional viral genomes.

If you work with biology content in Fable 5, what it refuses is substantially narrower now — though the highest-sensitivity categories still route to Opus 5 for additional scrutiny.

---

HTTP Terminator is a tool James Kettle built at PortSwigger — the company behind Burp Suite — to generate and evaluate HTTP attack hypotheses from RFC documents.

The tool ingested 138 RFC documents covering HTTP and SMTP specifications to produce its attack surface map. Across 30,000 authorized websites, it found approximately 700 vulnerable targets. Targets included banks and government infrastructure. A zero-day in Apache Traffic Server was among the findings and has since been patched. The vulnerability is tracked as CVE-2026-63078. The AI component proposed a novel vulnerability class: Shared-Parser Confusion, where servers using shared code for both request and response parsing expose response-processing features as attack vectors. Kettle validated the concept. HTTP Terminator was open-sourced and presented at Black Hat USA 2026. It uses Claude for document extraction and test-case generation.

If you build web services with shared request/response parsing logic, Shared-Parser Confusion is the class now named and documented — the PortSwigger Research publication has the technical specification.

---

If you're running parallel Claude Code sessions and have been routing coordination through shared files or external state, version 2.1.224 adds a more direct path: agents can now address messages to other named running sessions.

Two new tools enable it. ListAgents discovers reachable agents by name. SendMessage routes a text message to a named agent. Messages are text only — conversation history and files do not transfer between sessions. The feature is available on macOS and Linux. It is not available on Amazon Bedrock, Google Cloud's Agent Platform, or Microsoft Foundry.

If your multi-agent Claude Code workflow currently routes coordination through shared files, the direct messaging path is now available — with the constraint that managed cloud platforms are not yet in scope.

---