Anthropic AI Updates: October 13, 2026

1. Anthropic Details Distillation Defenses, Reports 10x to 100x Drop in Extractable Reasoning

Anthropic. Anthropic described the measures it has deployed against covert campaigns that extract Claude’s reasoning transcripts to train competing models, naming Zhipu (in connection with GLM-5.3) and Moonshot (Kimi) as alleged distillers operating through proxy networks of hundreds of thousands of fraudulent accounts. Defenses include “preserved thinking,” which stops API accounts created on or after August 31, 2026 from editing prior context, restricting Sonnet 5.5 reasoning transcripts to the requesting organization, and a two-stage classifier with a reported false positive rate below 0.1% running on Sonnet 5.5, Opus 5 and 5.5, and Fable 5 and 5.1. Anthropic reports that reasoning tokens extractable by distillers fell by 10x to more than 100x and that tens of thousands of linked accounts were banned, often within two hours of a campaign starting. For developers, the changes mean some integrations must be updated for preserved thinking, and since September 24 requests blocked by safeguards before Claude responds are billed. Source

2. Frontier Red Team Finds Small Open Models Approaching Self-Replication on IoT Devices

Anthropic. Anthropic’s Frontier Red Team published experiments suggesting small open-weight models may be months away from being able to self-replicate across insecure consumer devices such as Android TV boxes. In a simulated network of 8 emulated Android nodes with unauthenticated ADB ports, GLM-5.2 spread to all 8 nodes in 4 of 5 trials, while Qwen3.5-9B set up inference servers on 8 of 9 hardware types and Gemma4-26B-A4B ran across 15 TV boxes (about $55 each) at 0.39 tokens per second. The authors project, using the Epoch Capabilities Index trend for models of 30B parameters or fewer, that a small model could reach GLM-5.2’s level by the end of 2026, but note the exploit step was deliberately trivial and the spread would be slow, with doubling times of days. Recommended mitigations include ISP flow monitoring for multi-gigabyte model transfers, quarantine of infected subscribers, honeypots, and secure-by-default device configurations. Source