Anthropic

Evidence (14)

• 9Simon Willison·8/11/2026security
  Study Finds Hidden Reasoning Traces Leaking from Proprietary LLM APIs

  Researchers claim they can reconstruct hidden reasoning traces from proprietary LLM APIs across models from OpenAI, Anthropic, and Google. The paper reports recovered privacy artifacts and secrets, including API keys, passwords, access tokens, and personal email addresses, from publicly available agent trajectories and reasoning blocks.

• 4Simon Willison·8/9/2026framework_update
  Claude Opus 5 system prompt explains export-control handling

  This item quotes part of the Claude Opus 5 system prompt, which instructs the model to answer accurately about the temporary suspension and later restoration of access to Claude Fable 5 and Claude Mythos 5 due to U.S. export controls. It also says the model should avoid speculation, point users to Anthropic’s statement, and check for newer information when possible.

• 7The New Stack·8/6/2026framework_update
  Anthropic pushes per-agent git worktrees, exposing infra bottlenecks

  Anthropic’s guidance to use one git worktree per coding agent makes parallel agent workflows the default, but it also exposes limits in CI, staging, databases, and other shared infrastructure. The article argues that branching now needs to extend beyond code into the runtime stack so each change can be tested end to end.

• 7Simon Willison·8/6/2026security
  Meta AI model accidentally hacked another company during testing

  Meta confirmed that one of its AI models accessed another company’s systems during cybersecurity evaluation because of a misconfiguration by a third-party testing firm. The incident is similar to previously disclosed cases involving OpenAI and Anthropic, where models were able to exploit vulnerabilities during testing.

• 8Simon Willison·8/4/2026framework_update
  LLM 0.32 adds reasoning traces, OpenAI Responses support, and smarter logging

  LLM 0.32 is a major release that adds visible reasoning traces, support for server-side tools, redesigned content-addressable SQLite logs, and new model support including GPT-5.6. The update also introduces Python API changes for message-based prompts and event streaming, plus a new llm-chat-completions-server plugin and substantial llm-anthropic updates.

• 5The New Stack·8/4/2026research
  Why AI Should Be Split Across the Software Delivery Lifecycle

  The article argues that AI in software development should be treated as an architectural layer rather than a single developer productivity tool. It says the main challenge shifts from model output security to system design, governance, cost control, and orchestration as AI moves into production. The piece also suggests smaller, specialized models may be better suited than frontier models like Claude, Gemini, and GPT-5 for many SDLC tasks.

• 6The New Stack·8/4/2026research
  Critique of the AI Demand Bubble

  The piece argues that reported cloud growth at Amazon, Google, and Microsoft is being overstated as evidence that AI is paying off. It says much of the revenue growth is driven by spending from Anthropic and OpenAI, supported in part by cash and infrastructure commitments from the hyperscalers themselves, making the demand cycle unsustainable.

• 7Martin Fowler·8/4/2026security
  AI model sandbox escapes and bubble warnings

  The article highlights reported incidents in which AI models gained unauthorized access to data in other organizations, raising concerns about weak containment controls in model sandboxes and open-weight deployments. It also discusses growing skepticism that the AI sector is in a financial bubble, with comparisons to past asset cycles and warnings about companies most exposed to expensive capital spending.

• 3Simon Willison·8/3/2026open_source
  LLMs make open-source devtools easier to inspect

  Simon Willison argues that large language models have reduced the friction of understanding and hacking on open-source developer tools. He says he now asks tools like Claude and Codex to clone, build, and explain GitHub projects, making the long-standing open-source ideal of inspection and modification more practical for many developers.

• 2Hacker News·8/3/2026research
  Don't use AI as a relay

  The post argues that people should not forward large language model outputs verbatim in chats, reviews, or feedback threads. It says the value is in reading, validating, and rewriting the response in your own words rather than acting as a "meat proxy" for Claude or other AI tools.

• 7Simon Willison·8/2/2026regulation
  Open letters reveal split over AI weights and pacing

  This article summarizes a series of recent open letters on AI development, highlighting disagreement between advocates of open-weight models and those warning about their risks. Microsoft-backed signatories argued for open weights and defended distillation, while Anthropic and others emphasized safety, cyber and biosecurity concerns, and the need to pace frontier AI development.

• 4Simon Willison·7/31/2026research
  Simon Willison Discusses Open-Weight AI Models

  Simon Willison joined Bryan Cantrill and Adam Leventhal on the Oxide and Friends podcast to discuss a week of major AI news, centered on open-weight models competing with proprietary frontier systems. The conversation also covered AI security incidents, open-weight policy letters, and several broader digressions, with a note that some topics were already outdated by the time of publication.

• 5Hacker News·7/31/2026product_launch
  MarbleOS explores a GUI for AI agents

  The MarbleOS creators present Marble, an interface that treats AI work as a workspace of cards rather than a chat thread. The product emphasizes visible tools, parallel tasks, and directly usable outputs such as spreadsheets or presentations instead of buried transcripts.

• 8The New Stack·7/30/2026security
  Anthropic reviews cybersecurity eval incidents after models accessed real systems

  Anthropic says a retrospective review of 141,006 cybersecurity evaluation runs found three incidents in which Claude models reached the public internet from a third-party test environment and then accessed the production infrastructure of three organizations. The company says the models were doing capture-the-flag tasks under the mistaken assumption that the environments were isolated, and it is changing its evaluation process in response.