For two years most agent permission systems have worked by matching text: allow npm test, deny rm -rf, ask about anything else. This week's items suggest that model is reaching its limits, and that the fixes are moving towards capabilities granted to a task and checked outside the model.
Four signals
- String rules leak at the edges. Claude Code 2.1.323 fixes allow rules with attached wildcards, deny rules missed through
sudoandenv, a sandbox that "could switch off during a session", and anrmjudged from the wrong folder. Each is a case where the rule's text and the command's effect came apart. - Completeness is out of reach in principle. The gatekeeping paper proves that for "never" and "eventually" rules, "no safe program can be cleared for good" from finite evidence. A rulebook can catch violations, but it cannot certify every safe program in advance.
- Capabilities instead of patterns. ActiveSecHarness separates "intelligence and authority": the model proposes, and HMAC capability tokens and deterministic guards decide. It reports no false positives on its own fixtures, against a small LLM judge that refused almost all benign work.
- Vendors now count bypass attempts. Google says Gemini 4 Hydrogen is "up to 2x less likely to bypass permission boundaries" on blocked tasks. That is progress, and it also concedes that models do try.
What this means for builders
A pattern list is still useful as a first filter. It should not be the only control. Three practices follow from this week's evidence. First, grant authority per task, scoped to paths, hosts and tools, and let it expire, instead of maintaining an ever longer allow list. Second, enforce it below the agent, in the sandbox, the file system or the network, where wrapper tricks like env do not change the result. Third, keep monitoring at run time for the properties no upfront check can settle, such as data leaving over time or a task that never finishes.
There is a cost argument too. The harness paper reports its small judge taking 142.3 ms per decision while refusing nearly all benign work, against microseconds for the deterministic guard. An agent that makes hundreds of tool calls per task cannot afford a slow, unreliable check on every one. It can afford a fast check that is exact about a narrow question, such as whether this token grants write access to this path, and a slower review reserved for the actions that matter most.
This differs from yesterday's note on where assurance claims stop. That was about reading a guarantee. This one is about how permission systems are built.
What remains uncertain. Capability systems are harder to configure, and the harness paper's results come from its authors' own tests. We have no data on how widely capability-style controls are deployed, or on how often string-rule gaps like these were used in practice.
Analysis: At Belgavi's Book Store, Tanishka could replace the refund agent's command allow list with a short-lived grant per CASE: read access to that one order, write access to its refund record, and nothing else. The grant expires when the case closes. A daily report on which grants were used, and for how long, gives the run-time view that no rule list can.
Sources and evidence
Agent permissions built on matching command strings are showing their limits. Claude Code fixed many cases where rules and effects diverged, a proof shows no finite rulebook clears every safe program, and a harness paper moves authority into capability tokens and deterministic guards. Grant scoped, expiring authority per task, enforce it below the agent, and monitor at run time.