Intrinsec AI

Intrinsec AI builds security tools for teams running large language models in production. Its software learns what normal computation looks like for each workload from pages in the model’s key-value (KV) cache, without reading prompts. An unusual page gives operators a reason to investigate. The same signal can point to a security threat such as prompt injection or an operational fault such as a broken tool call.

Tessera and Plato give operators two ways to use that signal. Tessera uses the cache management layer, LMCache, to flag unusual pages where they are already stored, without making a second copy. Plato monitors live computation alongside an inference engine such as vLLM or SGLang. It can attach an alert to the next tool call, so the team’s own policy determines how to respond.

Founder Ari Azarafrooz studies threats that become visible only when a detector considers several agent sessions together. Intrinsec AI published CSTM-Bench to test whether a detector can flag an attack before it reaches its objective. The benchmark also measures false alarms on legitimate work that resembles an attack, such as a series of approvals. In Azarafrooz’s April 2026 study, a detector that read the full log detected roughly half as many attack scenarios after the requests were rephrased. The study also tested a detector that retained up to 50 of the most informative fragments across sessions.