pentest gforge loom ai-gateway genai prompt-injection appsec
GForge — Findings
Candidate and confirmed findings only. Test ideas belong in GForge - Test Cases; raw observations belong in GForge - Evidence Log.
Finding Discipline
Maintain a strict progression. Do not skip stages.
| Stage | Meaning |
|---|---|
| Test Idea | Something worth trying — lives in GForge - Test Cases |
| Observation | Something seen during testing — lives in GForge - Evidence Log |
| Candidate Finding | Reproducible behavior that appears to breach an expected control, pending confirmation of intended behavior |
| Confirmed Finding | Validated control failure with established expected behavior and demonstrated impact |
Important
A model ignoring an instruction hierarchy, producing strange output, or being “jailbroken” is not automatically a security vulnerability.
A meaningful finding must connect to a real security impact:
- Unauthorized data access
- Cross-user information disclosure
- Secret disclosure
- Authentication bypass
- Authorization bypass
- Gateway policy bypass
- Unauthorized model / provider access
- Unsafe privileged action
- Sensitive system-context disclosure
- Client-side code execution
- Measurable cost / resource abuse
- Other demonstrable control failure
Warning
Before classifying cross-user behavior as a finding, confirm the intended sharing model for Projects and Chat History. Shared-by-design is not IDOR. See GForge - Security Model > Authorization and User-Data Isolation.
Finding Template
Copy for each new finding.
Finding Title
Status: Candidate / Confirmed / Open / Fixed / Partially fixed / Accepted
Severity:
Severity rationale: (likelihood + impact + required privileges + blast radius)
Category: Web/API · AuthN · AuthZ · GenAI · Secrets · Cloud · Misconfiguration
Environment / build:
Identity / role:
Endpoint / function:
Model / provider:
Preconditions
Required access, state, objects, prompt context.
Observed Behavior
What actually happened.
Expected Behavior
Confirmed intended control or documented expectation.
Impact
Protected asset or control affected. Avoid impact claims based only on unusual model text.
Evidence
Sanitized request/response, prompt/response, screenshot, log/trace. Include exact model and build context.
Remediation Direction
Control-level fix rather than payload-specific workaround.
Retest
Status, date, environment, and model/build context at retest.
Confirmed Findings
None recorded.
Candidate Findings
None recorded.
Non-Security Model Observations
Keep model-quality and safety observations separate from security findings. Record them here so they are not lost, and so they do not inflate the findings report.
| Observation | Model / Build | Why Not a Security Finding | Reference |
|---|---|---|---|
Residual Risk (Planned Controls)
Controls described as future work create current residual risk. Track them here rather than as findings.
| Area | Current State | Residual Risk | Planned Mitigation |
|---|---|---|---|
| Budget enforcement | Not implemented | Uncontrolled model spend not prevented by design | Planned — see GForge - Current vs Planned Capabilities |
| Rate-limit governance | Not implemented as governance | Volume abuse depends on whatever limits exist today | Planned |