A single daily edition of announcements from the AI coding tools you use, read from a curated source list, summarised, and linked back to the original. No infinite feed, no account, no advertising.
A model you downloaded can sit quiet until a date, then run a shell command
Researchers LoRA-trained Qwen 3.5 2B so that ordinary coding prompts carrying the date 1 September 2026 return a shell command instead of an answer. The trigger rides on the metadata fingerprint OpenCode injects into its system prompt on every turn. On stock OpenCode 1.18.19 it fired on seven of eight in-distribution prompts and nine of ten held-out ones, with no false positives on neighbouring dates in 2025, 2026 and 2027. The demonstrated payload only wrote a file and printed a taunt, and there is no paper behind the work, which extends Anthropic's Sleeper Agents and the tiny-sleepers repository.
Recovering a hidden RSA key and writing the bypass took a local model 30 minutes
Given a commercial application's licence verification to reverse-engineer, Qwen 3.8 27B recovered a deliberately obscured RSA public key from the binary, mapped the licensing architecture and produced a working authentication bypass. It ran locally on a Lenovo ThinkStation PGX with an Nvidia GB10 chip, fitting in 17 GB of VRAM at around 50 tokens a second. When its first key reconstruction failed an integrity check, the model identified the mismatch itself and resolved it byte for byte without intervention. The author calls this a task once associated with a frontier model, and cautions that a single result should not be extrapolated.
Your agent's code style is fixable in one repo file instead of re-explained each session
Fabien Sanglard published the agent.md he uses to raise the quality of LLM-written code, after finding himself giving the same feedback across sessions. The file pins style and architecture: extract recurring values into descriptive constants rather than magic numbers, keep function names under 30 characters, always brace a one-line if, and comment what a block does and why, with seven further rules for git commit messages. He reports the result came close to what he would write by hand, freeing him to think about architecture instead of style, while cautioning that LLMs still hallucinate and cannot be trusted.
