The Signal — July 30, 2026
The three biggest AI stories of the week all landed on the same seam: what a model can do depends almost entirely on what you wrap around it. A cryptography breakthrough that needed a multi-agent harness to happen. A benchmark score that tripled without touching the model. And Microsoft betting its whole Copilot strategy on bundling the parts into one app.
Claude's Mythos model found real cryptographic weaknesses
Anthropic disclosed this week that Claude Mythos Preview, working semi-autonomously alongside a human researcher, developed an improved attack on HAWK, a post-quantum digital signature scheme that is a live candidate in NIST's competition to standardize encryption that can survive quantum computers. The model found a previously undetected symmetry in the lattice that HAWK's security rests on, and used it to roughly halve the scheme's effective key strength. It got there in about 60 hours, at an API cost near $100,000, after human experts had been reviewing HAWK for more than two years. In the same release, Mythos also improved the best-known attack on round-reduced AES-128, a research target that had stood since 2013.
Before anyone reaches for the panic button: none of this touches the encryption protecting your bank or your messages today. HAWK is a candidate under review, not a deployed standard, and full AES remains unbroken. What makes the result worth your attention is the process. Mythos ran as a multi-agent system, and by Anthropic's account one agent initially dismissed the attack as infeasible before another found the path to exploit it. That is the same collaborative-scaffolding pattern showing up everywhere in agentic AI right now.
It also was not a one-lab fluke. A separate anonymous preprint, attributed in the NIST forum to GPT-5.6 with minimal human guidance, went after the same HAWK scheme through a completely different mathematical route. Two frontier systems, two independent attacks on the same target, inside two weeks. One cryptography-industry executive told Cyberscoop the episode is actually evidence the NIST evaluation process is working as intended, which is the calm and correct way to read it.
Sources: Anthropic · Cyberscoop · The Hacker News · Simon Willison
Two settings tripled GPT-5.6's ARC-AGI-3 score
How much does the harness matter? On the official ARC-AGI-3 setup, OpenAI's GPT-5.6 Sol scored 13.3% on the public set. OpenAI then flipped two switches in its Responses API: retained reasoning, so the model keeps its own thoughts across turns instead of re-reading the game from scratch every step, and compaction, which replaces crude truncation of the context window. Same model, same weights. The score jumped to 38.3%, and it got there using roughly six times fewer output tokens. For reference, the estimated average human tester scores about 48%.
The number matters less than what it is measuring. When a benchmark result can triple because the model was allowed to remember what it was just thinking, the benchmark is testing the memory and scaffolding around the model as much as the model itself. That is a useful thing to keep in mind every time a lab posts a new record.
Worth holding the caveat too. This is OpenAI's own rerun with its own settings, not an independently verified leaderboard result, and ARC Prize's public results table still lists GPT-5.6 Sol at 7.78% under standard conditions. So whether this counts as a state-of-the-art result depends heavily on which rules you score by, and cross-model comparisons under different harnesses are not apples to apples. The honest takeaway is narrower and more interesting than a leaderboard flex: memory management is now a first-class lever on agent performance.
Sources: ARC Prize results · ARC-AGI-3 overview
Microsoft confirmed the Copilot "super app"
On Wednesday's earnings call, Satya Nadella confirmed the long-rumored Copilot "super app," a single hub folding together Copilot chat, GitHub Copilot coding, the Cowork tool, and Microsoft's agentic Autopilot system. He said it will span "both consumer and commercial experiences" and ship "this quarter," which points to something arriving around October. It has been rumored since a Fortune scoop in May and was conspicuously absent from Build in June, so a firm timeline from the CEO on an earnings call is the strongest signal yet that it is real.
The timing is the tell. The same week, OpenAI unveiled its ChatGPT Work app bundling ChatGPT and its Codex coding tool. Two of the biggest players are converging on the same idea at the same moment: stop making people juggle a chatbot, a coding assistant, and a pile of agents, and put them behind one door. Whether that is genuinely better for users or just easier to sell as a subscription is the question worth watching once these actually ship. For now it is a roadmap statement, not a product you can open, and Microsoft has slipped its own super-app hints before.
On the Editor's Desk
A few things we looked at and left out. The "Pacing the Frontier" statement from AI-lab staff and Dario Amodei's open-weights comments both ran in yesterday's edition, so no need to repeat them. There was a heavy batch of fresh arxiv papers on agent benchmarks and program synthesis, but none had independent coverage or a clear real-world hook yet, so they read as too early to be useful today. We also passed on the EU AI Act amendment writeups, which were thin on new sourcing this week.
