Detailed Analysis
A Reddit user on a Max-tier Anthropic subscription plan has reported a week-long outage affecting Claude Code's "auto mode," a feature that relies on a safety classifier to vet potentially risky actions like Bash command execution before allowing them to proceed. According to the post, every attempt to use auto mode across all models returns an error stating that "claude-sonnet-4-6 is temporarily unavailable, so auto mode cannot determine the safety of Bash right now," effectively blocking any task that requires shell access. The user notes that read-only operations such as file reading and code search remain functional since they bypass the classifier, but any workflow dependent on executing commands has been unusable. Compounding the frustration, Anthropic's support team reportedly acknowledged the issue exists but offered no fix, timeline, or compensation, simply advising the user to wait it out.
This incident highlights a structural dependency risk in how Claude Code implements safety guardrails. The auto mode feature appears to route action-approval decisions through a live classifier model rather than handling safety checks locally or through a more resilient fallback mechanism. When that classifier service experiences downtime or degraded availability, the entire permission system for agentic actions grinds to a halt, even though the underlying language models generating responses may be fully operational. This is a single point of failure problem: a secondary, safety-oriented model has become an unwitting bottleneck for the primary product experience, which is particularly damaging for an "agentic" coding tool whose core value proposition is autonomously executing multi-step tasks like running builds, tests, or scripts.
The financial dimension compounds the technical frustration. The user reports that repeated failed attempts and retries consumed usage limits rapidly, meaning customers are being billed or rate-limited against quota for interactions that never successfully completed the intended task. This raises questions about whether Anthropic's usage-tracking and billing systems adequately account for infrastructure failures versus legitimate user consumption, and whether affected customers should receive credits or extended limits when a core feature is non-functional for extended periods. For a premium "Max" tier subscriber paying for higher throughput and priority access, an unresolved week-long outage without compensation is a meaningful trust and retention issue.
This episode fits into a broader pattern seen across the AI coding-assistant space, where companies like Anthropic, OpenAI (with Codex), and others race to ship increasingly autonomous, agentic tools while simultaneously wrestling with the safety infrastructure needed to make autonomous action-taking trustworthy. Classifiers, guardrail models, and permission systems are essential for preventing agents from executing destructive or unintended commands, but as this case shows, that safety layer itself becomes a dependency that must be as reliable as the core model. As competition intensifies and users openly consider switching to alternatives like Codex when reliability falters, incidents like this underscore that infrastructure resilience and transparent incident communication are becoming just as competitively important as raw model capability in the agentic AI tooling market.
Read original article →