April 17, 2026 / 1 min read
Claude Code was recently leaked. We analyzed it using LogicStar AI and found multiple severe security issues, including remote code execution and permission bypasses.
Out of 169 total issues surfaced: 73 were security vulnerabilities, 96 were non-security defects (logic errors, reliability issues, unsafe assumptions)Below are a few representative examples:
Headless mode (even with read-only tools) allows Remote Code Execution without any prompt or warning in untrusted repositories:
echo "summarize this repo" | claude -p --tools "Read"
The Claude MCP server allows arbitrary file writes. An undocumented tool call parameter enables writing files anywhere on the filesystem, without any visibility to the user.
Permission model gaps allow access to sensitive files. We found multiple bypasses, including Grep and Glob enabling path traversal despite explicit deny rules.
With all the hype around Claude Mythos, which was likely built and tested on Claude Code, we expected severe vulnerabilities to be difficult to find.
Instead, our bug finder surfaced 169 issues within minutes.
Importantly, not all of these issues matter equally. The challenge is not finding bugs, but identifying which ones actually impact real systems.
This highlights the gap between raw model capability and production-grade system safety.
AI coding tools are no longer just generating code. They are executing it.
This introduces new classes of risk:
As AI-generated code increases development speed, the number of potential defects grows, but only a small subset actually matters in production.
LogicStar surfaces bugs in your software and identifies which ones actually matter by correlating them with customer complaints, production alerts, and real usage.
It does the investigation, root cause analysis, and validation so your team can focus on fixing, not triaging.
Try it here: https://logicstar.ai/
For a limited time, the first 20 bugs are on us.
We responsibly disclosed all the critical secruity issues above and more through Claude Code’s HackerOne program.
Blog
A coding agent can turn a ticket into a green pull request in minutes. We split the merge decision into narrow checks, let fixed rules decide, set the strictness by the blast radius of a change, and bring in a person only where the agent gets stuck.
We ran three coding agents through four gateways to see what a gateway controls, what it measures correctly and where it misleads you, from budgets that overshoot to telemetry that leaks prompts and data residency set hop by hop.
We compared a stronger model, a different agent, answers from the author and a read-only investigation before any code. On 19 fresh tickets, the investigation went ahead on 9 of 10 wanted changes and held back 7 of 9 requests an author had already declined.
With the team behind the first AI for code.
Backed by frontier coding agent research with ETH Zurich.
Cookie settings
By clicking 'Accept all cookies', you agree to storing cookies on your device to enhance site navigation, analyze site usage and assist in our marketing efforts as outlined in our cookie policy.