With #Polygraph, you can: -Expose latent biases: Move beyond surface outputs to measure what an LLM encodes as its true belief. -Contrast topics: Test whether a model encodes different internal stances on Topic A versus Topic B. - Directly compare how different LLMs represent

Takeaways: - The AI community still lacks reliable methods to evaluate and fix LLM failures. - Interpretability offers outsized impact - the main barrier to progress is that we don’t truly understand today’s models.

“𝐒𝐚𝐟𝐞𝐭𝐲 𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠” 𝐝𝐨𝐞𝐬 𝐧𝐨𝐭 𝐞𝐥𝐢𝐦𝐢𝐧𝐚𝐭𝐞 𝐛𝐢𝐚𝐬 𝐢𝐧 𝐋𝐋𝐌𝐬; it merely conditions models to suppress biased outputs under evaluation. Epos Labs introduces #AI #Polygraph. https://t.co/NiHCBx66ot

Superposition is the next buffer overflow

This means that a motivated attacker can abuse entanglement to undetectably manipulate LLMs. Nation State Actors are gearing up for the new opportunities an AI-powered software landscape will open for them:

What is Subliminal Learning? LLMs with several billion parameters are trying to represent the information contained in terabytes of web content. The math doesn’t check out - so instead LLMs cheat

Some takeaways for defenders: -You can’t rely on input-output filtering to detect attacks on models -You need to inspect your LLM supply chain -Subliminal attacks can be detected in real time

Without referencing the target behavior at all, the LLM finds itself with a high probability of performing the target action, due to a fundamental property of the neural network architecture.

Subliminal Learning Will Power the Next Generation of Influence Operations https://t.co/WHyKoJH665

Imagine an article about houseplants that causes AI to support Vladimir Putin. Bad actors use new attacks, turning AI into a weapon for disinformation and cyberattacks. See our demonstration of a Subliminal Attack here (and our "#Putinized" demo): https://t.co/WHyKoJH665