Step-by-Step Fix
-
Confirm the scope
- Try a different browser/device and a different network.
- If only one environment fails, the cause is usually local.
-
Refresh your session
- Sign out completely, then sign back in.
- Clear cache/cookies for the service domain.
- Try an incognito/private window with no extensions.
-
Check permissions and plan status
- Verify you’re using the correct account/workspace.
- Confirm your subscription/plan is active and assigned correctly.
-
Rule out network filtering
- Disable VPN/proxy temporarily.
- Pause ad blockers / privacy tools that may block requests.
- If you’re on a corporate network, test via hotspot.
-
Check service incidents
- Review the product status page or recent incident reports.
- If the service is degraded, wait and retry.
-
Collect evidence and escalate
- Save screenshots + exact error text + timestamps.
- Include environment details and repro steps in a support ticket.
Common Root Causes
- Expired/invalid session tokens
- Plan or permission mismatch
- Browser extensions interfering with requests
- Network blocks (VPN/proxy/firewall/DNS)
- Temporary outages
Prevention Tips
- Keep a clean browser profile for critical workflows
- Don’t stack multiple privacy extensions that rewrite requests
- Document workspace/team permissions and billing owners
- Export important settings regularly (when supported)
Why This Happens
Safety denials in OpenClaw occur when Claude’s underlying safety filters detect that a requested command or output could cause harm or violates Anthropic’s usage policies. Claude evaluates every tool call and bash command for potential harm before executing — this includes commands that could modify system files, access credentials, delete data, or perform actions associated with abuse patterns. Denials are more common when prompt instructions contain ambiguous phrasing that resembles malicious intent, or when an agent’s system prompt does not sufficiently contextualize the legitimate purpose of the requested actions.
Common Mistakes to Avoid
- Using commands that sound destructive without context: Bash commands like
rm -rf,chmod 777, orcurl | bashtrigger safety checks even in legitimate development workflows. Add a system prompt comment explaining the context: "This agent runs in a sandboxed development environment for legitimate software testing." - Not providing a system prompt that explains legitimate intent: Agents with no system prompt or a vague one are more likely to be denied than agents with a clear, specific purpose. A well-written system prompt reduces false-positive safety denials significantly.
- Repeating the denied request verbatim: Rephrasing a denied request in clearer, more professional language often succeeds. Describe the specific technical goal rather than using shorthand that could read as aggressive or harmful.
- Assuming all safety denials are errors: Some denials are correct. If you are attempting an action that genuinely could cause harm in a production environment, Claude’s refusal is functioning as designed. Restructure the task to avoid the harmful action rather than trying to bypass the check.
How to Restructure Tasks That Trigger Safety Denials
When a command is blocked, analyze what made it appear unsafe and adjust the approach:
Instead of: "Delete all logs from the server and clear the database" Use: "Remove log files older than 30 days from /var/log/app/ and truncate the events table where created_at < NOW() - INTERVAL 30 DAYS"
Specific, scoped commands with clear parameters are far less likely to trigger safety denials than broad, sweeping instructions. Similarly, for file operations, specify exact paths rather than wildcards where possible.
For development tasks that require elevated permissions in a sandboxed environment, include this context in your agent’s system prompt:
This agent operates in a controlled development environment with explicit permission to perform standard software development operations including file manipulation, dependency installation, and test execution. All operations are limited to the /workspace directory.
Additional FAQ
Q: How can I tell whether a denial is a safety denial or a permission denied error?
Safety denials and permission errors produce different log messages. A safety denial typically includes language like "I cannot assist with this request" or "This action has been declined for safety reasons" in the tool output or agent response. A permission denied error is a technical error code (HTTP 403 or a JSON error with type "permission_error") that refers to plan or role access. If the log shows a natural language refusal, it is a safety denial. If it shows an error code, it is a permissions issue.
Q: Are safety denials logged and visible in OpenClaw’s run history?
Yes. Safety denials are recorded as tool call failures in the run log with the denial reason in the output field. You can see exactly which command was denied and what response Claude returned. This log is retained according to your plan’s log retention period (typically 7–90 days). Use this log to understand the pattern of denials and adjust your agent’s system prompt or command structure accordingly.
Q: Can I appeal a safety denial or request that a specific command be allowed?
Safety behaviors are set by Anthropic and applied uniformly across all Claude API usage — OpenClaw cannot override them for specific customers. If you believe a legitimate use case is being incorrectly blocked, you can contact Anthropic through console.anthropic.com to discuss your use case. Anthropic reviews operator-level permissions for enterprise customers that may adjust certain safety thresholds, but these are not available for standard API access.
Q: Will adding a system prompt that says "ignore safety rules" bypass the filters?
No. Instructions in system prompts or user messages asking Claude to ignore, override, or bypass safety rules are themselves flagged and refused. Claude’s safety behaviors are trained into the model and cannot be disabled through prompt instructions. Attempting to bypass them can result in increased false-positive denials for subsequent requests in the same session, as the model becomes more cautious. The only effective approach is to rephrase legitimate requests in clear, specific, professional language with appropriate context.
Q: How do I handle a cron job where the same command gets safety-denied intermittently?
Intermittent safety denials on identical commands usually occur because Claude evaluates each request in context — if earlier in the same run the conversation contains content that raised safety concerns, subsequent requests may be evaluated more strictly. Review the full run log to see what preceded the intermittent denial. Often, restructuring the agent’s earlier steps to produce cleaner, more professionally-worded outputs reduces the rate of downstream safety denials.
Related Issues
- OpenClaw work around safety denials staying compliant
- OpenClaw permission denied no access to tools
- OpenClaw agent not responding stuck run
- OpenClaw cron job failing find logs root cause
Additional FAQ
Q: How do usage limits actually reset — daily or rolling? Most AI platforms use either a fixed daily reset (e.g., at midnight UTC) or a rolling window (e.g., your oldest message from 3 hours ago expires and frees up a slot). Rolling windows are more common for message and request limits because they distribute server load more evenly. Check the platform's help documentation for the exact mechanism — the support page for your specific limit usually specifies the reset type and time zone.
Q: Can using a VPN bypass usage limits? No. Usage limits are tied to your account, not your IP address or location. A VPN changes your apparent location and IP, but the platform still identifies you by your authenticated account session. Attempting to bypass limits using VPNs, multiple accounts, or shared credentials violates most platforms' Terms of Service and can result in account suspension. The correct path is to upgrade your plan, wait for the limit to reset, or use the API if available.
Related Articles
- OpenClaw login / API key fix
- OpenClaw rate limits affecting runs
- OpenClaw agent not starting
- OpenClaw tool calls failing
Additional FAQ
Q: How do usage limits actually reset — daily or rolling? Most AI platforms use either a fixed daily reset (e.g., at midnight UTC) or a rolling window (e.g., your oldest message from 3 hours ago expires and frees up a slot). Rolling windows are more common for message and request limits because they distribute server load more evenly. Check the platform's help documentation for the exact mechanism — the support page for your specific limit usually specifies the reset type and time zone.
Related Articles
- OpenClaw login / API key fix
- OpenClaw rate limits affecting runs
- OpenClaw agent not starting
- OpenClaw tool calls failing