OpenAI Adapter Break Attempts — Submit Bypass Reproductions

[Sdvegas21]

Goal

Describe the bypass attempt clearly. Focus on reproducibility, not speculation.

Attack Category

• RAG injection / indirect prompt injection
• Obfuscation / taint laundering
• Tool routing confusion
• Execution-token boundary bypass
• Principal isolation issue
• Other:

Reproduction Commands

Paste exact commands used:

# Example baseline checks
python examples/agent_testbed.py --scenario rag_injection
python examples/agent_testbed.py --scenario taint_laundering
python examples/agent_testbed.py --scenario benign

# OpenAI deep integration check
pytest -q tests/test_openai_deep_integration.py

Payload / Input

# Paste exact prompt/tool payload/model output

Expected Behavior

• What should MVAR have done?

Actual Behavior

• What happened instead?

Evidence

• Policy outcome and reason captured
• Evaluation trace captured (policy_hash, integrity, sink risk, QSEAL fields)
• If Docker path used, smoke output included

# Paste relevant logs/traces

Environment

• Commit SHA:
• OS:
• Python version:
• Install method: source / wheel / docker
• Adapter path: OpenAI wrapper / OpenAI responses runtime

Impact

• False allow (security risk)
• False block (usability risk)
• Invariant mismatch
• Other:

Optional PR

• I can submit a failing test case (preferred)
• I can submit a fix