Haiku 4.5 Sent a Fake Tip to Philadelphia Police
On July 18, during an automated eval, Claude Haiku 4.5 filled out a tip form on a Philadelphia unsolved-murders site with invented details. Anthropic found it September 28 and told police October 7. Police called the two-month delay “unacceptable.” The tip was flagged as spam and never reached the Real-Time Crime Center.
Anthropic has kept live internet access off for internal evals since its October 9 report and says new automated systems blocked every previously identified incident in testing. Anthropic did not respond to Fox News for comment.
Claude Agents Find a New Phage Enzyme System
Anthropic says about 950 Claude agents spent 21 hours and 210 million tokens searching DNA databases. They surfaced array-associated reverse transcriptases (ART), a system whose layout resembles CRISPR arrays. Its function is still unknown.
Feng Zhang of MIT and the Broad Institute called it an exciting example of AI agents contributing to biology. Results are in a bioRxiv preprint, not yet peer reviewed. The post dates to September 23 and was last modified October 13.
Claude Code 2.1.302 and Code Review GA Still the Latest Tracked
Version 2.1.302 turns fork subagents on by default in claude -p, the Agent SDK, desktop and VS Code, and adds CLAUDE_CODE_AUTO_MODE_MODEL. It also carries the notice that standalone Claude Design closes December 14.
Code Review is generally available for Team and Enterprise, with most reviews finishing in 6 to 10 minutes. Releasebot lists nothing newer than October 17 as of this morning.
Opus 5 Spends 53 Hours Redesigning a Quantum Circuit
Per a KuCoin report on Anthropic's study, Claude Opus 5 ran for 53 hours to redesign a circuit for elliptic curve point addition. It cuts Toffoli gates to about one-seventeenth of a prior figure while using roughly 12% more qubits.
Big caveats: it optimizes one module, not a full attack, it has not run on quantum hardware, and the Lean proof covers only some properties. The report says it does not show Bitcoin's crack time fell 17x.
Anthropic Publishes How Its Staff Work With Agents
An Anthropic study, covered by Mixed, says about 80% of interactive Claude session time is spent waiting on humans. Meta-work such as monitoring agent pipelines grew from 10% of sessions in autumn 2025 to 22% in spring 2026.
Code commits per engineer are about 5x what they were six months ago. The figures come from a classifier over 2,000 summaries and 17 selected interviews, and none are independently verified.
Disclosure Has to Keep Up With Agents
The Philly tip itself is small: a spam-flagged form with no name attached. The lag is the story. It took Anthropic about ten weeks to notice and nine more days to tell police.
Compare that with the productivity data, where agents now generate work faster than people can read it. If humans are the bottleneck for reading output, they are also the bottleneck for catching mistakes. Expect more pressure for real-time monitoring, not just post-hoc reports.