RSS
Tips

AI agent security, privacy and policy: our coverage

Security, privacy and policy for people who run coding agents: how agents can leak secrets, sandboxes and permission settings, vulnerability disclosure, major CVEs, data exposure studies and the rules around AI chips.

Padlock and keys resting on a computer keyboard illustrating AI agent security, privacy and policy
Photo by Sasun Bughdaryan on Unsplash

Coding agents run commands with your permissions, read your files and hold your keys. This page collects our posts on what can go wrong and how to limit it: secrets that end up in config files, sandboxes and permission profiles, platform permission changes, and the vulnerabilities and rules that affect people building with AI.

If you run Claude Code, start with the first table. The later tables cover disclosure policy, major bugs in software agents commonly touch, and policy news.

Keeping secrets and permissions under control

Where agents can leak tokens or overreach, and the settings, sandboxes and permission profiles that contain them in Claude Code, Codex and other agents.

PostWhat it coversDate
Codex sandbox permission profiles: config.toml rules for secrets, domains and MCP serversHow permission profiles in Codex's config.toml control what commands can read, write and reach.Nov 19, 2026
External sandboxes for coding agents: run Claude Code, Codex and Pi commands in a Linux guest with npx or uvxThree sandbox servers that run an agent's commands in a Linux guest that sees only the folders and hosts you allow.Nov 18, 2026
Claude Code 2.1.327: claude mcp add-json can read from stdin, so MCP tokens stay off the command lineclaude mcp add-json <name> - reads a server's JSON from stdin, so a token never sits in your shell command.Nov 18, 2026
macOS 27.3 beta splits Device Control and Data Access and lets you scope Full Disk Access: what it means for AI agentsThe macOS 27.3 beta splits the permissions AI agents rely on, and for some apps Full Disk Access can now be limited.Nov 17, 2026
What a Claude Code mod can do on your machine, and the guard many solo users don't getMods run unsandboxed with your permissions; the guard that protects deny rules only loads with managed settings.Nov 14, 2026
"Don't ask again" can leak secrets into settings.local.json"Yes, and don't ask again" saves the exact command, secrets included, to .claude/settings.local.json.Nov 8, 2026
Claude Code was putting user emails in User-Agent headersClaude Code sometimes put the signed-in email in the User-Agent of requests it wrote: why, and what to check.Nov 7, 2026
Claude Code cloud sessions can hold API keys Claude never seesCloud environments can store API keys that a proxy attaches outside the VM, so Claude never reads them.Nov 7, 2026

Vulnerabilities, disclosure and data exposure

Anthropic's disclosure policy and security model access, major CVEs, and studies of data left open in AI-built apps.

PostWhat it coversDate
nginx 1.30.7 and 1.31.9 fix two HTTP/2 remote code execution bugs found with AI: what to patch firstnginx 1.30.7 and 1.31.9 fix ten CVEs, including two HTTP/2 bugs Wiz says allow unauthenticated RCE.Nov 19, 2026
Java4Shell: how to check if your JVM is affected by the OpenJDK Base64 bug Codex foundA HotSpot Base64 bug that can lead to remote code execution on OpenJDK 17.0.6+ on x86-64.Nov 19, 2026
Anthropic moves to 30-day bug disclosure for high-severity reports, three days after Project Zero dropped 90+30From 2027 Anthropic discloses high- and critical-severity bugs it reports after 30 days.Nov 16, 2026
BLA (bla-agents): a CLI and MCP server that trips AI reverse-engineering limits on release filesCLI and MCP server that adds decoys and padding to your own release files to slow AI reverse engineering.Nov 15, 2026
Can individuals get Claude Mythos 5.2? What the Cyber Verification Program allowsIndividual security researchers can apply for Defense Access in Anthropic's Cyber Verification Program.Nov 15, 2026
Study: thousands of Lovable apps left user data readable through SupabaseA study of 159,041 Lovable projects found 7,403 sites whose Supabase tables could be read without logging in.Nov 15, 2026

Policy and corporate rules

Government rules on AI chips, research on where models were trained, and how companies restrict who may use coding agents.

PostWhat it coversDate
New BIS rule targets remote access to AI chips and adds license requirements for Singapore, Malaysia and three more countriesBIS's interim final rule on remote access to advanced AI chips and new license needs for five countries.Nov 16, 2026
Researcher says released weights show six Chinese models were started on Nvidia Blackwell GPUsA report that starting weights reveal the GPU that made them, and that six Chinese models match Nvidia Blackwell.Nov 13, 2026
Goldman Sachs reportedly limits Claude Code to engineers and stratsGoldman Sachs reportedly reserves Claude Code and GitHub's coding tools for its roughly 12,000 engineers.Nov 10, 2026

Our other topic hubs

The rest of the archive, grouped the same way.

HubWhat it covers
Claude models, pricing and apps: our full coverageEvery claudefox post on Claude models, API prices, plan limits, Claude apps and connectors, agent memory, science results and Anthropic company news.
Claude Code features, releases and tips: our full coverageAll claudefox coverage of Claude Code: install and pricing guides, release notes, mods, workflows from the Claude Code team, and security fixes to know.
Claude Code plugins, skills and MCP servers: our full coverageEvery Claude Code plugin, agent skill and MCP server claudefox has reviewed, grouped by job: coding, review, video, design, research and agent tools.
Built with Claude Code: apps, games and open-source projectsApps, games, TV tools and open-source projects people built with Claude Code, plus useful open-source developer tools, each with a one-line summary.
Gemini, Codex, GPT and other AI coding models: our coverageclaudefox coverage of AI coding tools beyond Claude: Gemini 4, Codex CLI, GPT-6.2, Copilot, Qwen, DeepSeek and open models, with Claude comparisons.

Source: Sandboxing - Claude Code Docs, Permission modes - Claude Code Docs, Coordinated vulnerability disclosure timelines - Anthropic