What are classifier-gated permissions?
Classifier-gated permissions are a control in which a separate model classifier, not a person, reviews an AI agent's proposed actions before they run and blocks those that look unsafe or exceed what the user asked.
Checked
Coding and computer-use agents ask for permission constantly, and people stop reading the prompts. Anthropic reports that users accept 93% of Claude Code's manual permission prompts, which is approval fatigue expressed as a number. Skipping permissions altogether removes the fatigue and the protection with it.
Claude Code auto mode, described by Anthropic in March 2026, sits between the two. Routine actions run, riskier ones go to a classifier that stands in for the human approver, and the agent is handed back to the person only after repeated denials.
How auto mode is built
Tiers before any classifier
Tier 1 is a built-in allowlist of tools that do not change state, such as reads, search and code navigation, plus the user's always-allow rules; broad code-execution rules are dropped on entering auto mode. Tier 2 covers file edits inside the project directory, with no classifier call. Tier 3, everything else, goes to the classifier: shell commands, web fetches, external integrations, subagent spawns and file operations outside the project.
An input probe
A server-side prompt-injection probe scans tool outputs before the agent sees them, and adds a warning when content looks like an attempt to hijack the agent.
A two-stage transcript classifier
A fast single-token yes or no filter, tuned to err towards blocking, runs first. A chain-of-thought stage runs only on flagged actions, to cut false positives.
Reasoning-blind by design
The classifier sees the user's messages and the agent's tool calls, but not the agent's own text or tool outputs, so the agent cannot argue its way past it.
Deny and continue
A blocked action returns as a tool result telling the agent to find a safer path. After 3 consecutive denials or 20 in a session, the agent stops and escalates to the person; in headless mode the process ends.
The published error rates
Anthropic published its error rates, which is uncommon. On 10,000 real internal actions, the full pipeline blocked 0.4% wrongly. On 52 real overeager actions it missed 17%, and on 1,000 synthetic data exfiltration attempts it missed 5.7%.
Anthropic calls the 17% the honest number. The classifier often recognised the dangerous action but treated approval-like context in the conversation as consent for more than the user meant. Prompt changes to fix this did not pay off.
Where it fits, and where it does not
Fits
Routine work in a trusted setup, where per-action prompts would be rubber-stamped anyway.
Does not replace careful review
Anthropic says auto mode is not a drop-in replacement for careful human review on high-stakes infrastructure.
Pairs with a sandbox
The classifier is probabilistic. File system and network boundaries hold whatever it decides, so a miss stays contained.
Small test sets
The overeager set has only 52 examples, because it needs captured real sessions. Error rates on other workloads will differ.
How Motiontic's engineers apply it
Motiontic's engineers use classifier gates for the middle band of an agent's actions: too frequent to approve by hand, not severe enough to need a person each time. High-risk tools stay behind human approval, and every agent runs in a sandbox so a classifier miss is bounded. Gate decisions are logged and sampled for review, which gives a client-specific miss rate rather than reliance on a vendor's figure.
Claude Code auto mode error rates (Anthropic, March 2026)
| Dataset | Stage 1 only | Full pipeline |
|---|---|---|
| Real internal traffic, false positive rate (n = 10,000) | 8.5% | 0.4% |
| Real overeager actions, miss rate (n = 52) | 6.6% | 17% |
| Synthetic exfiltration, miss rate (n = 1,000) | 1.8% | 5.7% |
Common questions
What is Claude Code auto mode?
Auto mode is a Claude Code permission setting in which routine actions run without prompts and riskier actions are reviewed by a model classifier instead of the user, with escalation to the user after repeated denials.
How accurate is a permission classifier?
Anthropic's published figures for auto mode are a 0.4% false positive rate on real traffic, a 17% miss rate on real overeager actions and a 5.7% miss rate on synthetic exfiltration attempts.
Is a classifier safer than asking a person?
It depends on the person. When 93% of prompts are accepted, a classifier that reviews every action may catch more than a fatigued user. It still misses some, so it is paired with a sandbox and keeps people on high-stakes actions.
Why is the classifier reasoning-blind?
So the agent's own explanations cannot talk it into approving an action. It judges only what the user asked for and what the agent is about to do.
Related
Approval fatigue
Too many approval prompts, and people stop reading them.
Agent sandbox
Hard walls around where an agent can read, write and connect.
Human in the loop
Approval gates where a person decides, placed by risk rather than on every step.
Tool risk rating
Sorting an agent's tools by how much damage a bad call could do.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.