Agents hooked into GitHub can steal creds – but Anthropic, Google, and Microsoft haven't warned users

Security researchers hijacked three popular AI agents that integrate with GitHub Actions by using a new type of prompt injection attack to steal API keys and access tokens, and the vendors who run agents didn’t disclose the problem. The researchers targeted Anthropic's Claude Code Security Review, Google's Gemini CLI Action, …

[Alan Mackenzie]

100, 500, 1337???

So that's how many dollars researchers get paid for uncovering critical security vulnerabilities. This is truly pathetic. You could increase these sums by a factor of 100, and they wouldn't be unreasonably high.

The researchers could surely make much more by selling to the dark side. It's likely this has already happened lots of times.

[IGotOut]

Re: 100, 500, 1337???

But it's AI, that's like a nano second loss. They gotta stem the flow somehow.

[druck]

"The solution?"

"The solution?" - disable this agent shit, or use a git sever that doesn't support it.

These security failures can never be fixed due to the very nature of LLMs, and allowing them access to system commands is like playing Russian Roulette with a semi-automatic.

[Anonymous Coward]

Re: "The solution?"

Why access tokens and API key are allowed anywhere near a repository like this is nuts.

Anything like this should be via integration to a User/password/credential/token/key management tool.

[[e-mail]]

Duh Magazine

Isn't this pretty obvious? Anyone who hasn't considered this shouldn't be allowed near a keyboard.

[Bebu sa Ware]

The Reg image for this story is too good to miss

Manglement ?

[MrBill]

I wonder if Mythos would detect these vulnerabilities

But I don't think so.