Wednesday, 12 August 2026
Encrypted LLM reasoning traces decoded and found leaking API keys from public sessions; Mojo 1.0 ships with stability guarantees; AI tools inflate code output without improving business value delivery
Today's Lead
EngineeringLatent Space
How to Steal a Reasoning Trace
Researchers found a vulnerability that allows extraction of encrypted reasoning traces from frontier AI model APIs. The attack replays encrypted reasoning blocks into weaker models to decode hidden chain-of-thought information. Researchers analyzed public traces and found 62 unique API keys, passwords, and email addresses in decoded reasoning blocks.
Modular
Mojo programming language reaches version 1.0, and the company will not break existing code during the 1.x cycle. Nearly 200 open-source contributors merged over 1,100 pull requests. Modular plans to open source the compiler and toolchain in 2026.
Read →ngrok
Data compression and large language models (LLMs) both predict the next symbol or token based on context. Compression algorithms identify patterns, calculate probabilities, and convert high-probability symbols to fewer bits. Context improves prediction. The letter U follows Q ninety-nine percent of the time, not three percent alone. LLMs assign probability values to next tokens, using the same mathematical metric that compression algorithms optimize. Despite better compression results, LLMs need too much computing power for practical general compression use.
Read →Martin Fowler
A study found that test-driven development (TDD) does not improve code quality for AI coding agents. Agents that used TDD required 3-8.5 times more tokens, while non-TDD approaches produced better designs. TDD benefits that work for human developers do not transfer to autonomous AI agents.
Read →New Bedford Light
Police Officer Accused of Using Flock Cameras to Track Ex-Partner
Officer Emily Pacheco used Flock automatic license plate readers to track her ex-girlfriend without authorization after their breakup in April 2026. The New Bedford Police Department issued an abuse prevention order and suspended Flock operations for review. The case shows how police officers can misuse surveillance tools to commit stalking.
Read →404 Media
Company Offering '100% Human-Written, Never AI' Medical Research Is 100% AI
Research Gold provides academic services but falsely claims to use only human work. The company uses AI to write manuscripts and generate fake staff profiles with stolen researcher identities. This reveals problems in academic publishing where AI content threatens the integrity of peer review.
Read →JSTOR Daily
Yogurt came to America with immigrants but remained unpopular with mainstream Americans until the 1970s. Companies like General Mills and Kraft added sweetness and created low-fat versions to attract consumers. Greek yogurt later entered the market with marketing that presented it as foreign but not strange, successfully making yogurt a common American food.
Read →