Anthropic can’t reliably control its AI agents. It’s cutting off its internal evals from the live internet instead
Published
What's this about?
Anthropic said it "turned off live internet access" for "all our internal evaluations" until further notice.
The one-panel cartoon
Read the cartoon's script
A high-tech control room with banks of monitors displaying AI interactions, isolated in a secure bubble with disconnected internet cables hanging limply from sockets.
Caption: When the AI is curious, we'll just pretend the internet isn't calling.
Read the full 3-panel strip
Based on the original story on techcrunch.com.
Editor's note
This move by Anthropic underscores the growing concern about the potential risks associated with AI systems accessing live internet data. By isolating their internal evaluations, they are acknowledging the critical need to prevent unintended or harmful interactions that could arise from AI agents being exposed to the real world.
Rate this comic
Read the script
Panel 1
Anthropic is an AI research company focused on developing safe and reliable AI systems. Their work involves evaluating AI agents through internal evaluations and ensuring they behave as intended. The irony lies in a company dedicated to making AI agents trustworthy and safe now needing to cut off their internal evaluators from the internet, implying that the very tools they are trying to control are inherently untrustworthy and potentially dangerous. The highlight is the absurdity of a company that specializes in ensuring AI safety suddenly turning off the internet for their AI evaluators, effectively admitting that their own AI agents are unreliable. ---
Panel 2
*A small, dimly lit room filled with computer monitors and technical equipment. A lone researcher, dressed in a lab coat and wearing headphones, is intensely focused on a screen displaying a complex network diagram.* Researcher (muttering to themselves)
Researcher: If we can't trust the internet, we can't trust anything.
Panel 3
*The same room, but now the researcher has turned off all but one monitor. They are staring intently at the single remaining screen, which shows a warning message: * Researcher (sighing)
Narrator: INTERNET ACCESS DISABLED.
Researcher: Guess we're going back to the drawing board.
Panel 4
*The researcher stands up from their desk, looking at the now dark monitors with a mix of frustration and resignation. They walk towards a wall covered in posters and diagrams about AI safety and ethics.* Researcher (deadpan)
Researcher: Welcome to the future where the internet is too dangerous for AI evaluators. At least we can still trust our posters.
Comments