The Intake
The Intake — Wednesday, August 12, 2026
On the substrate
Frontier AI reasoning blocks shared a decryption key across model variants; researchers extracted credentials from 315,320 public samples
Panfilov et al., arXiv 2608.09867 Simon Willison
If you've been treating encrypted reasoning traces from frontier AI APIs as private to your model, session, or user account, a paper published August 10 names what made that assumption fragile. Within a provider's model family, all variants shared a single decryption key. Panfilov et al. demonstrated the attack by feeding a Claude Opus 4 reasoning block into a Claude Haiku 4.5 prompt. Haiku decoded and returned the Opus reasoning as plaintext. Claude Haiku 4.5 was the most vulnerable model in the disclosed testing.
The team analyzed 315,320 publicly scraped reasoning blocks. They extracted 367 PII artifacts and 182 credentials. The paper names four attack vectors: anti-distillation bypass, private data extraction, hazardous information disclosure, and invisible prompt injection. All three providers acknowledged the report. Each has since patched the vulnerability.
If you've built products that surface reasoning traces to end users, the key-sharing vulnerability is patched.
Claude Code auto mode becomes the default on August 14 for Pro, Max, and Team plans
If you're running Claude Code on a Pro, Max, or Team plan, your default interaction mode changes on August 14. Auto mode allows Claude Code to proceed without requesting approval unless an action is classified as irreversible, destructive, or aimed outside your environment. The change follows a safety study with 1,053 paid users. In that study, Anthropic says the classifier caught 89% of deliberately inserted dangerous commands. Human reviewers caught 13.6%. The same study found that users approved 97% of permission prompts without genuine scrutiny.
A session-level fallback activates if auto mode blocks three consecutive actions. A second threshold triggers at 20 total blocked actions within a session. When either limit is reached, Claude Code reverts to manual approval for the rest of that session. If you're on an Enterprise, API, or cloud platform deployment, auto mode starts as opt-in rather than the default.
---
For operators
Three experimental Anthropic API endpoints and the legacy Workbench retire August 17
If your workflow calls any of Anthropic's experimental prompt-tool endpoints, the cutoff is five days away. Three endpoints go offline August 17, 2026: /v1/experimental/generate_prompt, /v1/experimental/improve_prompt, and /v1/experimental/templatize_prompt. Calls to these endpoints return an error after that date. The legacy Workbench at platform.claude.com/workbench also loses access on August 17. Anthropic announced the deprecation July 17, giving 31 days of notice.
Anthropic Sonnet 5 introductory pricing is now the permanent standard rate
If you've been building cost models with Claude Sonnet 5 expecting a September 1 rate increase, Anthropic confirmed August 10 that it won't happen. The introductory rate is now permanent: $2 per million input tokens. Output stays at $10 per million tokens. The previously announced increase to $3 per million input and $15 per million output will not occur. If you've priced in that rate change, you can revert those projections.
---