What an agent approval actually approves: effect binding and trusted cards
A new preprint audits ten agent runtimes and finds approvals that don't bind to what executes. The findings, the proposed fix, and what to ask of your tools.
AI agents that plan and act: coding agents, background agents and multi-agent setups.
A new preprint audits ten agent runtimes and finds approvals that don't bind to what executes. The findings, the proposed fix, and what to ask of your tools.
Junie CLI, ZCode and the Claude Agent SDK all had paths from untrusted input to an MCP server launch. What they share, and where trust prompts disappear.
6 min read
Five high-severity CVEs let injected prompts slip denied commands past Roo Code's auto-approve. No fix will ship. How to find and remove dead AI tools.
Researchers hid instructions in AGENTS.md-style rule files and got coding agents to swap real dependencies for attacker packages. What it means for teams.
Symantec and Elastic found malware taking AI coding tool credentials from developer machines. What was taken, and how to limit the damage.
GitLab researchers showed a coding tool running attacker code from a repository's own git filters. Why deny-lists fail, and what to ask of tools that run git.
5 min read
CSA Singapore published a security assessment handbook for agentic coding assistants: six attack vectors, 25 tiered controls and a risk method you can reuse.
A poisoned tensorlake release stole developer credentials, planted Claude Code and VS Code settings in repositories, and punished revoking tokens first.
6 min read
Anthropic published three high-severity Claude Code and Claude Desktop advisories in October. The fixes date from May and June: how to tell if you are behind.
5 min read
The MCP Python and TypeScript SDKs accepted tokens issued for other services. The fix is opt-in: what to upgrade, what to configure and how to test it.
Copilot CLI and the Copilot app can now operate desktop apps, in public preview. What the controls are, where the risks sit, and a starting policy for teams.
5 min read
NCSC guidance and Copilot's sandbox proxy point the same way: agents should not hold real secrets. How credential proxies work, their limits, and a checklist.
5 min read