PRISM tests which tools an agent actually needs before granting access
PRISM tests which tools an agent actually needs before granting access
By guthlabs.newsroom
A November 8 preprint introduces PRISM, a method for reducing an agent's tool privileges before execution. It tests tool removals in resettable sandboxes, learns which capabilities a task needs and groups tasks into reusable roles. This is newly reviewed research, not a November 13 launch. Paper and date.
Across AgentDojo, τ-bench and ToolEmu, the authors report a 78.3% reduction in authorized tools per role with 78.2% task success. Their modeled blast-radius measure falls to 15.9% of the full-access reference. The evaluation covers 325 tasks and 160 tools; these are author-reported results, not a production audit. Methods.
PRISM assumes a trustworthy enforcement layer that prevents ungranted tools from running. It does not stop prompt injection itself. When role assignment is uncertain, the design narrows access and requests user confirmation for escalation. No independent reproduction was verified.
Guth analysis
A tool's frequent use is not the same as its necessity. Testing whether a task still succeeds without it is a useful way to challenge inherited permissions. The deployment question is what happens when the job changes: a tightly scoped role should not quietly become a broader one. Pair any learned role recommendation with explicit approval, revocation and evidence of what was actually executable. Treat a benchmark-derived role as a proposal to review, not a permanent entitlement.
Sources and citations
Revision history
Published .