posts · 2026-08-12

This one's worth opening because it dismantles the "encrypted reasoning stops distillation" line that model vendors have been pushing. The first paper shows encrypted reasoning blocks from Anthropic, OpenAI, and Google are interchangeable across models—attackers can spend $720 using cheap Haiku 4.5 to decode Opus 4.8's real reasoning traces in bulk. The second goes further: without touching encrypted blocks, an inversion model trained on a 1.5B weak model can reconstruct GPT-5.4 mini's reasoning from public outputs alone, lifting a student model's MATH500 accuracy from 68.4% to 76.0%. I'd discount this a bit since both are still arXiv preprints, but the experimental design looks solid. The security part is more immediately concerning. Researchers decrypted 6,708 public Agent traces from GitHub and found 62 API keys, 33 passwords, and 7 private keys—64 of these secrets never appeared in the plaintext conversation. Developers sharing session logs for debugging can't inspect or scrub these opaque blocks. Vendors encrypted reasoning to protect their own IP, but the review and cleanup burden lands entirely on API users. Don't panic yet—this only affects models with encrypted reasoning enabled—but if you're building long-running Agents and reusing others' traces, it's time to check your logs.

I'd open this because it maps "AI-done" onto five concrete levels, which is way more honest than a green checkmark. OpenAI dropped 10 math results on Aug 1 with Lean 4 proofs—8,666 build tasks, zero errors. Five days later the paper went from 249 to 253 pages to patch a gap in an edge case. The code didn't change, so the machine only verified syntax and logic rules, not whether the proof actually matched the original problem. Terence Tao proposed a new priority rule: credit goes to the first team that delivers the full package—paper, explanation, and formal certificate—not just the code. I read this as an engineering signal for agent workflows: passing unit tests is only level two, don't treat it as the finish line. Only one of the ten results has reached level five, "absorbed by the community"; most are stuck at levels two or three. The post doesn't name the five unreviewed results, but the split is already clear.

OpenRouter ran Exa, Parallel, Perplexity, and native lab search engines across four benchmarks—BrowseComp, DeepSearchQA, WideSearch, HLE—with different models and search turn budgets, then published live leaderboards. The headline finding: search budget is the single biggest quality lever. On BrowseComp, Claude Opus 5 with Perplexity went from 35.8% at 1 turn to 89.0% at 25 turns, while cost only rose 2.5–7×. On easier tasks like HLE, extra turns barely moved the needle—GPT-5.6 Sol scored similarly at 1 and 25 turns but cost 3× more. The catch: when models can't find an answer, they burn through the full budget, making worst-case costs spike. The leaderboards update live; OpenRouter recommends testing against your own workload. I'd focus on the BrowseComp cost curves for hard fact-finding tasks—extra turns on easy benchmarks just waste money.