[BUG] Recurring false-positive Usage Policy refusal triggered by screenshot/vision content, not prompt text

Open 💬 0 comments Opened Jul 4, 2026 by shubhsaxena2020

Summary

Recurring false-positive "Usage Policy" refusals in Claude Code, consistently triggered by vision (screenshot) input rather than any text prompt content. Confirmed 12 distinct refusals in a single day (2026-07-04) across 9 unrelated sessions/tasks.

Environment

Claude Code version: 2.1.200 / 2.1.201
Entrypoint: claude-vscode (VS Code extension), also reproduced via CLI
Platform: Windows 11
Auth: Claude subscription (claude.ai OAuth) — confirmed no third-party gateway/proxy in the request path for any of these sessions

Pattern observed

In every case I traced, the refusal immediately follows an MCP tool call that returns a screenshot of a legitimate small-business website under active development. The refusal fires regardless of the actual task being performed that turn — tasks included pricing-card redesign, a git/deploy push, CSS/video fixes, and general dev-server review, none of which are themselves policy-sensitive. The same screenshot content (not the coding task) appears to be the common factor across all 12 occurrences.

As documented in your own error reference (https://code.claude.com/docs/en/errors#usage-policy-refusal), the check evaluates the full conversation, so retrying in the same session (or via --continue/--resume) reliably re-triggers the same refusal, since the flagged image remains in the transcript. This matches what was observed, including one case where the user explicitly pushed back that the content wasn't policy-violating, and the retry failed identically.

Request IDs (2026-07-04, chronological, UTC)

00:17:36 — req_011Ccg4DRbPd1BbAdhWSu3xv
00:19:58 — req_011Ccg4Po5zhjdJkQXuuDjY3
01:07:05 — req_011Ccg7zguVo9PhoBqHnpeQz
10:30:49 — req_011CcgrzZkPo9iuYD4g7j8UL
10:32:33 — req_011Ccgs7m3z28khUDSEneoJx
10:44:47 — req_011Ccgt49nia7jPxudRXerz1
10:53:34 — req_011CcgtiqZCDtbV6EDBbC9ZP
11:02:23 — req_011CcguQBvNv19biUvyy8drM
11:48:27 — req_011CcgxuZredPLLVpYX87s8K
12:21:28 — req_011Cch1SE4Nf3fu2aF31KLdK
17:09:03 — req_011CchPMcpGdTwfmJwDygkmc
17:20:14 — req_011CchQDF2EgYJNZuTesJUQ2

I'm happy to share the specific image content privately with an Anthropic engineer if useful for tuning the classifier — I'd rather not post it in a public issue.

Expected behavior

Legitimate business-website screenshots shouldn't repeatedly trip the Usage Policy classifier across unrelated development tasks in one day.

Actual behavior

12 refusals in one day, all following the same screenshot pattern, forcing session rewinds/restarts to continue unrelated work.

View original on GitHub ↗