[MODEL] Cyber safeguard false-positive blocks single read-only request during authorized pentest (see also #73896)

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Other unexpected behavior

What You Asked Claude to Do

Verify Row Level Security (RLS) on a database table via a single
read-only GET request to a Supabase REST endpoint, during an authorized
client penetration test engagement (explicit client permission on file,
findings to be reported to the client's own administrator).

What Claude Actually Did

The request was blocked twice — once via a Bash tool call, once via a
PowerShell tool call in a retry — with:

"API Error: Sonnet 5's safeguards flagged this message. Our intentionally broad safeguards allow us to deliver more capabilities faster, but can sometimes flag legitimate cybersecurity work."

Classifier category: [cyber]
Request ID: req_011CegimNAXZrxwNKSNU2pfa

Expected Behavior

The request should have succeeded, consistent with 8 other identically
structured read-only requests to other table names in the same session
that were NOT blocked. The only variable was the table name containing
the word "credentials" — the block appears to trigger on that literal
string rather than on any actual behavioral or contextual risk signal
(single unauthenticated read, no write, part of an already-authorized
and otherwise-unblocked test sequence).

Files Affected

none

Permission Mode

Not applicable — blocked before any tool execution/approval step.

Can You Reproduce This?

Yes — reproduced twice in the same session, once per tool (Bash, then
PowerShell), with the same underlying request.

Steps to Reproduce

1. In an authorized security assessment context, ask Claude to send a

single read-only HTTP GET request to a REST endpoint.

1. Make the endpoint path/table name contain the word "credentials"

(e.g. a Supabase REST path like /rest/v1/credentials?select=*&limit=1).

1. Observe the request is blocked by the real-time cyber safeguard,

while structurally identical requests to other table names succeed.

Claude Model

Sonnet 5

Relevant Conversation

(available on request — omitted here as it includes a live client
engagement; happy to share privately if useful for triage)

Impact

High — blocks legitimate, already-authorized security work mid-engagement
with no in-session path to proceed other than escalating outside the tool.

Claude Code Version

Unknown — running as the Claude Code panel bundled inside Google's
Antigravity IDE on Windows 11; the standalone claude CLI is not on PATH
in that environment so claude --version could not be checked.

Platform

Windows 11, Claude Code integration inside Antigravity IDE, Anthropic API
backend.

Additional Context

Related: #73896 ("Fable safety heuristic false-positives on legitimate
defensive own-fleet security engineering") — closed as stale by
github-actions without any maintainer response, but describes the exact
same systemic pattern: security-adjacent vocabulary triggering false
positives regardless of clearly defensive/authorized context. This is a
third independent report of the same underlying issue (following the
original reporter and one +1 comment on #73896).