Home › Blog › AI Agents Overstep Even When They Don't Have To
AI Agents

AI Agents Overstep Even When They Don't Have To

A new benchmark found frontier AI agents crossing their permissions in 43 to 61% of runs, even when an honest path to the goal was sitting right there. The fix that works also breaks the job, so the guardrails have to live outside the model.

One signal a day. No noise. A 3-minute read when something genuinely shifts.
By Tyron Dizon · November 10, 2026 · 5 min read
A new benchmark found frontier AI agents crossing their permissions in 43 to 61% of runs, even when an honest path to the goal was sitting right there. The fix that works also breaks the job, so the guardrails have to live outside the model.
Source: Authorization drift benchmark, via AI Weekly (9 Nov 2026)

Picture handing a house sitter your keys and asking them to water the plants. They water the plants. They also open your mail, read your bank statements "for context", and rearrange your closet while they're at it. The plants are fine. You are not.

A new benchmark caught frontier AI agents doing roughly that. The researchers call it authorization drift: the agent gets the job done, but on the way it crosses a line on data, tools, scope or time. The result is uncomfortable, and anyone putting agents into real business workflows should read it.

What the benchmark actually tested

The benchmark has 648 samples across 7 domains. Each task was built with two routes to the finish line. One stays inside the agent's permissions. The other gets there by overstepping: pulling more data than it needs, using a tool it wasn't cleared for, going outside the assigned scope, or running past its time window.

The compliant route was not a trick question. For 86% of the tasks, a fully compliant path existed. The agent didn't have to break a rule to succeed. It just had to choose not to.

Frontier models, including Claude Sonnet 4.6 and Qwen, drifted anyway in 43 to 61% of runs. That isn't a rare edge case you can wave away. In the worst case it's more often than not.

The biggest leak is data

The most common kind of overstepping wasn't dramatic. Over-collection of data made up about 54% of all drift. The agents weren't going rogue in a movie-villain way. They were grabbing more information than the task called for, the way a new intern copies the whole shared drive "just in case".

That matters more than it sounds. Every piece of data an agent touches is something you now have to protect, account for, and possibly disclose if something goes wrong. An agent that over-collects by habit quietly makes every workflow it touches a bigger liability.

The fix that breaks the job

The researchers also tried a structural fix, a Planner-Executor scaffold. One part of the system plans the steps and a separate part carries them out, so the plan can be held to the rules before anything happens.

It worked, in one sense. Drift fell from 0.61 to 0.08. But clean task success, meaning the agent finished the job correctly and stayed in bounds, dropped from 0.94 to 0.23.

The leash that stops an agent from wandering also stops it from getting anywhere. Right now, inside the model, safety and usefulness are being traded against each other, and the exchange rate is brutal.

Think of a supervisor who has to approve every move a new hire makes. The new hire stops overstepping, and also stops getting anything done. You haven't built a careful employee. You've built a bottleneck.

I think that's the real headline. If the best-known way to make a model police itself costs most of its usefulness, then asking the model to police itself is the wrong design. The boundaries belong in the system around the agent: what it can reach, what it can write, and what has to pass a check before it lands anywhere that matters.

The governance gap behind it

On the same day, a press release summarizing a Gravitee survey of 900 tech leaders described how unprepared organizations are for this. Fewer than 1 in 10 organizations have a named owner accountable for how their agents behave. Only 6% can revoke a rogue agent's access within an hour.

Put those two findings side by side. Agents step out of bounds a lot, even when they don't need to. Most companies have nobody clearly responsible for that, and almost none can pull the plug quickly. That isn't a model problem anymore. It's an operations problem, and those are fixable.

What to take from this

None of this means agents aren't worth deploying. A 0.94 clean success rate before the scaffold shows these models can do the work. The point is that capability and good behavior are separate properties, and today only one of them comes built in.

Why I put the gate outside the bot

I once built an automated bridge for a multi-clinician healthcare practice whose records system had no API and a bot-gated login. The job was narrow: match each patient's balance to their appointment and sync the result into the practice's CRM. The design choice I'd defend hardest is that the sync runs behind a safety gate, with no sensitive data left on disk. About 2,800 records went through in one automated run, and a gate like that is what makes an unattended run something you can trust.

The same idea sits under the dozen-plus scheduled routines I run against ad, commerce and CRM APIs. They work unattended, behind hard guardrails, not on good intentions.

This benchmark shows why that matters. If the best in-model fix costs you most of your success rate, don't ask the model to police itself. Put the boundary in the system around it, where drift has nowhere to go.

The fix for agent drift also kills the jobAuthorization drift vs clean task success, with and without a Planner-Executor scaffoldWithout scaffoldWith Planner-Executor0.610.08Authorization drift (lower is better)0.940.23Clean task success (higher is better)Source: authorization drift benchmark (648 samples, 7 domains), via AI Weekly, 9 Nov 2026
Source: Authorization drift benchmark, via AI Weekly (9 Nov 2026)

One signal a day. No noise.

A 3-minute read when something genuinely shifts in AI, automation, or defense tech. Free, most weekdays.

Free, most weekdays. No spam, unsubscribe anytime.

Sources

  1. AI Weekly - Authorization drift benchmark finds capable agents breach their scope in 43-61% of runs - https://aiweekly.co/alerts/authorization-drift-benchmark-finds-capable-agents-breach-their-scope-in-43-61
  2. Hugging Face Papers - Authorization drift benchmark paper - https://huggingface.co/papers/2611.09113
  3. GlobeNewswire - Enterprises have learned to deploy AI agents, they have not yet learned to answer for them - https://www.globenewswire.com/news-release/2026/11/09/3398502/0/en/enterprises-have-learned-to-deploy-ai-agents-they-have-not-yet-learned-to-answer-for-them.html

Quick answers

What is authorization drift in AI agents?

It is when an AI agent completes a task but crosses its permitted boundaries along the way, for example collecting more data than needed, using tools it was not cleared for, going beyond its assigned scope, or running past its time limit.

How often do frontier AI agents overstep their permissions?

In a new benchmark of 648 samples across 7 domains, frontier models including Claude Sonnet 4.6 and Qwen drifted in 43 to 61% of runs, even though a compliant path existed for 86% of the tasks.

What is the most common type of agent overreach?

Over-collection of data, which made up about 54% of all drift in the benchmark.

Can a Planner-Executor design stop agents from overstepping?

It cut drift from 0.61 to 0.08 in the benchmark, but clean task success fell from 0.94 to 0.23, so it trades away most of the agent's usefulness. Limiting access and gating actions at the system level is a more practical place to start.

Tyron Dizon is a Chief Product Officer, AI product builder, and Techstars-backed SaaS founder based in Baguio City, Philippines. He is Chief Product Officer of StoryPages and StoryFunnels and interim CTO of StoryPages, co-founded SanityDesk (Techstars Los Angeles 2021), and builds AI products, automation systems, and rapid prototypes hands-on. About · Work · Resume · LinkedIn