Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-10 · high
Generating the proof takes about 43 times longer than verifying it, deciding where in an agent's loop it can run.
Why now: It's new work, tested only against one benchmark so far, with no report yet of production use.
NiyamAI wraps every AI agent tool call in a cryptographic proof, per a paper posted to arXiv as 2608.07167.
Teams giving an agent access to tools have to decide how much to trust an automated approval. NiyamAI's numbers give them one data point: on Agent-SafetyBench, it catches unsafe tool calls with an F1 score of 88.5% at a 1.1% false-positive rate.
Before a tool call runs, the system checks it against a SHA-256-locked set of rules called an Intent Contract. A separate, isolated model acts as judge. NiyamAI then generates a zk-SNARK proof, via EZKL, that the check happened, without revealing anything about the underlying model's weights.
The catch is speed. Generating each proof takes 2260.6 milliseconds on average. Verifying one takes 53.1 milliseconds, roughly 43 times faster. Two and a quarter seconds is a tolerable tax on a single high-stakes, irreversible tool call. It's a dealbreaker for an agent calling tools one after another in a loop.
Each link below shares sources, entities, or timing with this story.
Shared entities / Same source domain / Earlier coverage
Both cover Agent, SHA; reported by the same outlet (arxiv.org); earlier Agent coverage from 2026-06-26.
Shared entity: Agent / Same source domain / Shared topic / Earlier coverage
Both cover Agent; reported by the same outlet (arxiv.org); overlapping topics (action, against).
Both cover Agent; reported by the same outlet (arxiv.org); overlapping topics (action, cost).
Same source domain / Shared topic / Tension
Reported by the same outlet (arxiv.org); overlapping topics (against, call, check, tool); pushes against this story (but).
Reported by the same outlet (arxiv.org); overlapping topics (against, call, deployment, tool); pushes against this story (against).
Shared entity: Agent / Same source domain / Earlier coverage / Tension
Both cover Agent; reported by the same outlet (arxiv.org); earlier Agent coverage from 2026-08-04.
Shared entity: Agent / Shared topic / Earlier coverage / Tension
Both cover Agent; overlapping topics (against, call); earlier Agent coverage from 2026-08-03.
Shared entity: Judge / Shared topic / Earlier coverage / Tension
Both cover Judge; overlapping topics (against, call); earlier Judge coverage from 2026-08-02.