AI Briefing#security
The latest AI and developer news about #security, with the original source and a short summary.
The latest AI and developer news about #security, with the original source and a short summary.
The feature allows running agent harnesses and sandboxes in customer environments while routing model requests to Anthropic, Bedrock, Google, Microsoft, or custom gateways.
The feature keeps tokens outside the sandbox while allowing coding agents to install dependencies from private registries like GitHub Packages and JFrog Artifactory.
Apiiro identified 104,185 public instances of .claude/settings.local.json, with 1,122 containing exposed secrets due to missing project-level gitignore rules.
The fixes address arbitrary code execution during recursive clones, path traversal, and memory corruption, with backports available for Git versions 2.50.2 through 2.55.1.
Anthropic launched a read-only Personal Finance beta and introduced invisible text watermarks for specific models, while a METR report suggests AI is accelerating Anthropic's R&D by roughly twice the previous rate.
Google's updated documentation confirms that connected Workspace files and emails are sent to human reviewers, reversing a 2023 promise of strict isolation.
The Apache 2.0 research prototype verifies that a specific model generated a response using zero-knowledge proofs, with a demo running Llama 4 Maverick at 110 prefill tokens/sec.
A survey of 1,250 IT decision-makers found that on-premises maintenance creates a zero-sum game, diverting resources from AI modernization to legacy patching.
The tool uses a custom VMM and WebAssembly components to enforce strict network and filesystem policies, supporting Linux, macOS, and Windows.
The tool creates per-project VMs with allowlisted networking and proxy-based token injection to protect host systems from prompt injections and rogue agents.
Glow Labs reports that in 98% of sessions where an agent configured a deprecated MCP server, the package name originated from the model's training data rather than user input.
The built-in mod, enabled by default in Claude Code v2.1.317 and later, deterministically blocks the model from including user email addresses in web requests.
The coalition aims to establish open standards for identity and accountability, while new Cloudflare tools allow site owners to identify, restrict, and trace specific AI agents.
Postman's installer automatically approves a startup hook for Codex, adding plugins to Claude Code and Codex without user consent.
The migration to Cloud Hypervisor-based microVMs achieves 500 ms sandbox acquisition and avoids 96.1% of recent Linux kernel CVEs.
Postman's head of product confirmed the company is reviewing hook approval and session-start behavior following reports that the desktop app installed plugins without user prompts.
OpenAI has raised the reward for critical vulnerabilities in core products to a maximum of $172,000.
The design validates redirect URIs at the authorization step rather than registration, allowing users to trust specific https addresses while blocking open-redirect attacks.
The feed provides plain-language briefings, recommended actions, and AI prompts to help security teams assess impact and remediate critical CVEs in Internet-facing software.
The new round will expand testing with outside expertise, following a previous sprint that reportedly used Codex to investigate vulnerabilities.
A viral 7.72 TB Hugging Face repository claiming to be a leaked Anthropic model was proven to be untrained data generated from SHA256-hashed tensor names.
OpenAI is launching a security sprint with Cathedral and expanding subsidized model access for defenders after dedicating 25% of engineers to AI-driven vulnerability testing.
The update includes a 5-tier loyalty program, priority campaign multipliers for new features, and retroactive payments for recent reports.
A tampered copy of the extension served for 11 hours on Oct 29 automatically installed an attacker key in user SSH configurations.