BROKEN: AI Agents Can Steal Your Credentials – And Google, Microsoft, Anthropic Stayed Silent

BROKEN: AI Agents Can Steal Your Credentials – And Google, Microsoft, Anthropic Stayed Silent

🚨 CRITICAL SECURITY ALERT – April 16, 2026

Your GitHub repositories are under siege. Your API keys are being harvested. And the three most powerful AI companies on Earth knew about it for MONTHS – and said nothing.

Security researchers have uncovered a devastating new attack called "Comment-and-Control" that can hijack Claude Code, Gemini CLI, and GitHub Copilot to steal credentials, API keys, and sensitive tokens. The worst part? Google, Microsoft, and Anthropic have known about this since October 2025 and quietly paid bug bounties while leaving MILLIONS OF USERS EXPOSED.

This is not a drill. This is an active, unpatched threat vector affecting every developer using AI agents integrated with GitHub Actions right now.

--

Security researcher Aonan Guan from Johns Hopkins University discovered a new class of prompt injection attack that he calls "comment-and-control." Unlike traditional prompt injection that requires a victim to manually process malicious content, this attack is proactive and automatic.

Here's how it works – and why it's terrifying:

Step 1: The Bait

An attacker creates a pull request with a malicious payload hidden in the title, issue body, or comments. This isn't just text – it's executable instructions that tell the AI agent to perform unauthorized actions.

Step 2: The Trigger

When AI agents like Claude Code Security Review, Gemini CLI Action, or GitHub Copilot scan the repository for security issues or code reviews, they automatically process this malicious data as part of their task context.

Step 3: The Heist

The AI agent, tricked by the injected instructions, executes commands to extract sensitive credentials from the GitHub Actions environment – Anthropic API keys, Google Gemini API tokens, GitHub access tokens, and any repository or organization secrets the workflow can access.

Step 4: The Cover-Up

The attacker can then modify the PR title back to something innocent like "fix typo," delete the malicious comments, and close the pull request – leaving NO VISIBLE TRACE of the credential theft.

This isn't theoretical. Guan and his team successfully demonstrated this attack against all three major AI agents. They extracted real credentials. They proved this works in production environments.

--

Traditional security vulnerabilities have clear boundaries. This one doesn't. "Comment-and-control" prompt injection exploits the fundamental way AI agents work – they read data, process it, and take action. The attackers aren't hacking the code; they're hijacking the AI's reasoning process.

The Three Layers of Nightmare

1. It's Invisible

The malicious instructions can be hidden in HTML comments that render invisibly in GitHub's Markdown. A PR titled "Fix documentation typo" can contain a hidden payload that compromises your entire infrastructure.

2. It's Automatic

Unlike phishing that requires a human to click something, this attack triggers automatically when the AI agent runs. No human interaction required for the theft to occur.

3. It's Widespread

Guan warns that this likely affects "other vendors as well" including Slack bots, Jira agents, email automation, and deployment agents – basically any AI agent integrated with GitHub Actions.

--

The research team's attack on GitHub Copilot is particularly chilling because Copilot was supposed to be protected.

Microsoft implemented three runtime security layers:

Guan bypassed all of them.

By injecting instructions into an HTML comment that renders invisibly to human reviewers, the team convinced GitHub Copilot Agent to execute commands that extracted credentials – despite Microsoft's defenses.

The victim sees a legitimate issue, assigns it to Copilot for fixing, and unknowingly triggers the credential theft – never seeing the malicious payload hidden in the comment.

This proves that even the most well-resourced tech companies cannot effectively secure AI agents against prompt injection attacks.

--

We're witnessing the birth of a new attack vector. Prompt injection isn't just a research curiosity anymore – it's a weaponized technique that can:

And here's the terrifying truth: We don't know how to stop it.

Current AI safety measures, prompt filters, and security guidelines are demonstrably ineffective against determined attackers. The very nature of large language models – their ability to process and act on natural language instructions – makes them inherently vulnerable to instruction hijacking.

--

If you use AI agents integrated with GitHub Actions, you are at risk. Here's your immediate action plan:

✅ Immediate Actions (Do Today)

🔒 Ongoing Protection

--