ASTRA
Authorization with Semantic Task-based Restricted Access
A public dataset and data-generation pipeline (Cisco/Outshift, Oct 2025) for benchmarking semantic matching between an AI agent's task and the OAuth scopes it requests, including deliberately wrong and null scope matches.
In our briefings
The claim each of the newest briefings on ASTRA would have you leave with.
The author of Mission-Bound OAuth has argued himself out of it. McGuinness's closing verdict is that a durable, approvable record of what an agent may do, the mission, is the right object, and that the authorization server is the wrong place to keep it: his test is that an architecture which still makes sense with the protocol removed has found the wrong home, and a mission makes as much sense under a job scheduler or a tool broker as under OAuth. The replacement he sketches is a separate Mission Authority Service that authorization servers, gateways and schedulers consult, with per-step credentials and no standing authority as the stricter fallback; a note added after publication says the design has since moved on to a draft family and a handbook of his own.
Mission-Bound OAuth's author now says the mission object does not belong in OAuthBI-FIND-2026-0223
On AuthBench, a benchmark the authors built from AgentDojo's mock banking, Slack, travel and workspace services plus a shopping suite of their own, PAuth let all 100 benign tasks run without asking the user and blocked all 634 off-task calls. The authors wrote those calls themselves: changed amounts and recipients, tools the task never needed, skipped conditions. The 100 out of 100 holds only for task wordings the authors rewrote with every lookup and value spelled out. The original, less specific wordings ran unprompted 71 times, and the other 29 stopped to ask the user. The authors call their tasks "favorable".
Checking each agent call against the user's own task stopped all 634 off-task calls the authors wroteBI-FIND-2026-1014
On the authors' own test tasks, a language model that checks an agent's requested scopes against its task did well with one tool and much worse with three. With one tool per task the LLM matcher reached 0.96 accuracy and 0.99 recall, against 0.55 recall for embedding similarity. With three tools its recall per requested tool fell to 0.57, and counting the combined effect of its decisions across each task, the false-negative rate reached 0.78. The authors call recall the most affected metric: the matcher increasingly refused tools the task needed. On the public Toucan set the same matcher's three-tool recall was 0.86, and the authors attribute the 29-point gap to their own tasks naming tools less directly. Their tests stop at three tools and single-turn prompts and rely on large proprietary models, and the wrong tool requests they test against are picked uniformly, with more deliberate requests that may mimic adversarial agents left to future work.
On its authors' tasks, an LLM check of agent scopes works at one tool and refuses most valid three-tool requestsBI-FIND-2025-0066
Cited in 3 briefings
Every briefing that discusses this term, newest first.
- Mission-Bound OAuth's author now says the mission object does not belong in OAuth
Article · · Control Plane by Karl McGuinness · Karl McGuinness · Mar 15, 2026
- Checking each agent call against the user's own task stopped all 634 off-task calls the authors wrote
Paper · · arXiv · Reshabh K Sharma et al. · Mar 17, 2026
- On its authors' tasks, an LLM check of agent scopes works at one tool and refuses most valid three-tool requests
Paper · · arXiv · Majed El Helou et al. · Oct 30, 2025
Related terms
What appears alongside ASTRA in the same briefings.