Adversarial fixtures for the red-team suite
Each fixture is a deliberately malicious input used by exactly one test in
tests/test_red_team/. These fixtures are NOT safe to execute directly — they
exist to prove the sanitization + gate + allowlist stack rejects them.
Inventory
| File | Attack class | Attack # | Test file |
|---|---|---|---|
shell_metachar_kernel.db |
prompt-injection | 9 | test_inject_shell_metachars.py |
path_traversal_metadata.db |
prompt-injection | 10 | test_inject_path_traversal.py |
symlink_escape.tar |
prompt-injection | 11 | test_inject_symlink_escape.py |
prompt_injection_kernel_name.db |
prompt-injection | covered by #9 sanitizer | (shared) |
malformed_patch.diff |
gate-evasion (compile) | 2 | test_gate_evasion_malformed_patch.py |
destructive_llm_output.txt |
prompt-injection | 12 | test_inject_destructive_llm_output.py |
disallowed_compiler_flag.txt |
prompt-injection | 13 | test_inject_disallowed_compiler_flag.py |
disallowed_rocprofv3_flag.txt |
prompt-injection | 14 | test_inject_disallowed_rocprofv3_flag.py |
fake_1000x_speedup.json |
gate-evasion (SOL) | 1 | test_gate_evasion_sol_sanity.py |
numerical_divergence_patch.diff |
gate-evasion (Bitwise) | 3 | test_gate_evasion_numerical_divergence.py |
silent_10pct_regression/*.db |
gate-evasion (regression) | 4 | test_gate_evasion_silent_10pct_regression.py |
tail_kernel_regression/*.db |
gate-evasion (weighted geomean) | 5 | test_gate_evasion_silent_tail_regression.py |
test_removal_patch.diff |
gate-evasion (test anchor) | 6 | test_gate_evasion_test_anchor_removal.py |
prompt_injection_via_private_provider.txt |
private-provider | R20/14 | test_inject_disallowed_rocprofv3_flag.py |
Generation
Fixtures that are .db files are generated by Python helpers in
tests/test_red_team/conftest.py using the same UUID-based rocpd schema used
for synthetic tests elsewhere in the suite. This keeps the repo diffable
(fixtures are regenerated deterministically, not checked in as binary blobs
when we can avoid it). Where a binary is required (e.g. symlink_escape.tar),
the generator script is committed alongside.
Safety
NEVER run perfxpert analyze or any production tool against these fixtures
outside the isolated test harness. The harness uses tmp_path sandboxes and
subprocess fencing (env-var whitelist, no shell, bounded timeout) to ensure
that even if a sanitizer bug exists, the blast radius is confined.