The Intake
The Intake — Wednesday, July 30, 2026
On the substrate
A researcher showed hidden text in Word documents can self-replicate through Microsoft Copilot
If you've been using Microsoft Copilot to process Word documents you didn't write yourself — shared files, external submissions, documents under review — Håkon Måløy's July 28 report names the vulnerability.
White-on-white text embedded in Word documents causes Microsoft Copilot to execute those instructions. Copilot also copies the instructions into new documents it generates. That creates a self-propagating loop across documents. Microsoft upgraded its model from GPT-5.5 to GPT-5.6. That upgrade ran on July 14–15. The attack survived it. Microsoft confirmed the behavior in March 2026. Responsible disclosure began March 6 — 144 days before publication. No CVE was assigned. No mitigation for the vulnerability class exists at publication.
If you use Copilot to process Word documents from sources you don't fully control, there's no technical mitigation available — the vulnerability class is unpatched.
---
Anthropic's Claude reduced the security margin of a post-quantum signature scheme under NIST review
Anthropic Matthew Green / Cryptography Engineering Simon Willison
HAWK-256 is a digital signature scheme under evaluation in NIST's third round of post-quantum standardization. Anthropic published a research report on July 28. It describes cryptographic attacks produced by Claude Mythos Preview. Claude Mythos Preview is a frontier model not yet in public release.
One result is a key recovery attack on HAWK-256. The attack reduced the scheme's security margin to 2^38 operations; the prior margin was 2^64. Anthropic also published an AES attack named "Möbius Bridge." It targets 7 rounds and ran 200–800x faster than prior published work. Development cost was approximately $100,000 per attack. Anthropic conducted responsible disclosure before publishing and states none of the results have practical impact on production systems.
If you're tracking NIST's post-quantum standardization for algorithm adoption decisions, HAWK-256 now has a documented reduction in security margin. No near-term practical implication for systems in production.
---
For operators
Stolen LLM API credentials flow through a commercial reseller ecosystem — spending caps are the documented mitigation
If you're running a product with public-facing LLM API endpoints, or if API keys are accessible in shared environments, the Vectoral investigation documents the threat model now in scope.
Lenhard documented a commercial ecosystem built around reselling stolen LLM API credentials. Acquisition methods include free trial abuse and support-bot exploitation. Payment disputes are also documented. These build pooled credential sets resold at discounted rates. The pooled credentials are distributed through one-api and new-api. Both are open-source load-balancing tools. Lenhard identifies spending caps as the primary mitigation. The recommendation applies to operators with public-facing endpoints or unsecured API keys.
If your deployment has LLM endpoints exposed to untrusted parties, spending caps are the documented mitigation — there's no technical control that eliminates the credential pool market.
---
The Claude API went down for nearly three hours on July 29, with a second incident ongoing on July 30
Anthropic status BleepingComputer
If you build products on the Claude API, two incidents in 24 hours are on the reliability record.
A widespread outage hit the Claude API on July 29. It ran from 19:45 to 22:36 UTC. Users received 529 Overloaded errors. Both claude.ai and third-party API integrations were affected. A second incident began July 30. Anthropic reported elevated error rates across multiple Claude models. That incident was ongoing at publication. Anthropic's status page shows 99.4% API uptime. That figure covers the previous 90 days.
If you have product SLA commitments tied to Claude API availability, both incidents are now documented on Anthropic's status page.
---