- #12 — 2026-W23Tool-Augmented AI Agents Often Learn the Protocol, Not the Capability2 pieces · 1724 words
- #11 — 2026-W22LLM Agents Break More on Paraphrases Than on Reformatting1 pieces · 1765 words
- #10 — 2026-W21LLMs Ignore Tool Access Rules Up to 68% of the Time1 pieces · 3593 words
- #9 — 2026-W20A Single Misleading Document Can Tank Long-Context AI Performance1 pieces · 2646 words
- #8 — 2026-W19AI Code Generators Build Working Software That Rots From Within1 pieces · 1732 words
- #7 — 2026-W18Frontier AI Models Don't Sabotage Safety Research — Yet1 pieces · 2853 words
- #6 — 2026-W17Web Coding Benchmarks Finally Test What Matters: Visuals, Interaction, and Repair1 pieces · 4605 words
- #5 — 2026-W16Why Reverting an AI Agent's Instructions Doesn't Undo Its Behavior1 pieces · 2729 words
- #4 — 2026-W12Visual Inputs Break Moral Safety Filters in Vision-Language Models1 pieces · 2722 words
- #3 — 2026-W11LLM Agents Can Now Post-Train Other LLMs — With Caveats1 pieces · 3568 words
- #2 — 2026-W10.1Safety Alignment Backfires in Non-English Languages Across LLM Groups1 pieces · 2800 words