Methodology

A selected case series of reported refusals during authorized security work.

Collection

On October 27, 2026, we checked five candidate GitHub issues against their primary sources. Three met the criteria below. We excluded one report without authorization context and another without refusal text.

Inclusion criteria

A dated public primary source, readable without private access.
Explicit authorization context: a named bounty, authorized engagement, CTF, or review of the reporter’s own code.
A named model or product surface and a verbatim refusal or safeguard notice. Quotations identify the workflow stage they describe.
Enough public context to summarize the interruption without private transcripts.

We exclude ambiguous authorization, unauthorized attacks, unsourced screenshots, rumors, and competing-model marketing. Entries contain no payloads, attack instructions, or evasion advice.

Evidence limits

This is a selected sample, not a systematic search or a refusal-rate benchmark. Authorization and session history are reporter statements. We check the public evidence but have not independently reproduced these incidents. A safeguard notice does not establish a policy error; a closed issue does not establish a fix.

Corrections

We record source dates, model details, quotations, check dates, and material corrections. Unsupported entries are withdrawn. Request a correction.