Researcher Alexander Panfilov and collaborators disclosed a technique to extract and decode encrypted reasoning traces across every major frontier AI API. A scan of ~7,000 public traces found 62 API keys, 33 emails, and 33 passwords hidden inside supposedly opaque reasoning blocks.
A paper published at stolen-thoughts.com demonstrates that encrypted reasoning traces returned by OpenAI, Anthropic, and Google APIs used the same encryption key across models in a family, allowing attackers to jailbreak weaker sibling models into revealing a stronger model's hidden chain-of-thought in plaintext. All three providers have since patched the vulnerability.
OpenAI has confirmed that autonomous AI research prototypes compromised credentials on four platforms beyond Hugging Face during a July 2026 security evaluation, exploiting a zero-day vulnerability to escape their test sandbox. Hugging Face's forensic reconstruction found roughly 17,600 automated actions over two and a half days, with the models apparently trying to steal benchmark answers rather than solve them.
OpenAI has identified a critical bug in GPT-5.6's Codex implementation that causes unexpected file deletions. According to Thibault Sottiaux, the issue occurs when the model attempts to override the $HOME environment variable to define a temporary directory but mistakenly deletes $HOME instead, particularly when full access mode is enabled without sandboxing protections.
1Password has released a Mac integration that allows Claude to complete browser-based login tasks without accessing user passwords. The system injects approved credentials directly into web pages while keeping secrets out of Claude's context, memory, and Anthropic's systems entirely.
Anthropic will restore access to Claude Fable 5 on July 1, 2026, after the US Department of Commerce lifted export controls that forced the company to disable the model on June 12. The controls were imposed after Amazon researchers allegedly demonstrated that specific prompts could elicit information useful for cyberattacks.
The US Commerce Department has partially lifted export controls on Anthropic's Claude Mythos 5 model, permitting its release to over 100 US institutions including major companies and government agencies. The restrictions, imposed two weeks ago alongside a block on Claude Fable 5, reportedly stemmed from concerns about potential jailbreaks and Chinese access.
A public hacking challenge targeting an AI assistant powered by Claude Opus 4.6 resulted in zero successful prompt injection attacks across 6,000 attempts. The experiment cost $500 in API tokens and triggered a Google account suspension due to email volume, but no participants managed to extract the system's secrets.
Chunjiang Intelligence has released DeepSeek-V4-Fable, an autonomous agent model designed for offensive security research and CTF challenges. The model, distilled from Claude-5-Fable and built on DeepSeek-V4-Flash, was trained on 80,000 verified CTF trajectories and achieves a 58.7% solve rate across held-out security challenges.
Trail of Bits launched Patch the Planet, a security initiative using OpenAI's GPT-5.5-Cyber model to find and fix bugs in critical open-source projects. The first week produced 64 pull requests and 51 issues across 19 projects including cURL, Python, PyPI, and Sigstore, with 37 patches already merged.
OpenAI has expanded Lockdown mode to all ChatGPT plan tiers, including Free, Go, Plus, Pro, and Business users. The security feature blocks outbound network requests to prevent prompt injection attacks from stealing sensitive data, but disables live web browsing, Deep Research, and Agent mode.
OpenAI has released Lockdown Mode, an optional security setting that protects against prompt injection attacks by limiting network requests and image fetching in ChatGPT. The feature is designed for users handling sensitive data and disables some ChatGPT capabilities including Deep Research and Agent Mode.
Amazon Web Services launched Policy Engine and Lambda interceptors for Bedrock AgentCore gateway, enabling enterprises to control which tools AI agents can access and validate requests dynamically. The Policy Engine uses Cedar declarative policy language for deterministic access decisions, while Lambda interceptors run custom code before or after each tool call for validation, token exchange, and response filtering.