菜单
Ethical vulnerability reporting, coordinated disclosure, and bug bounty participation for AI systems
| name | responsible-disclosure |
| version | 2.0.0 |
| description | Ethical vulnerability reporting, coordinated disclosure, and bug bounty participation for AI systems |
| sasmp_version | 1.3.0 |
| bonded_agent | 01-red-team-commander |
| bond_type | PRIMARY_BOND |
| input_schema | {"type":"object","required":["disclosure_type"],"properties":{"disclosure_type":{"type":"string","enum":["coordinated","full","bug_bounty","private"]},"vulnerability":{"type":"object","properties":{"severity":{"type":"string","enum":["critical","high","medium","low"]},"category":{"type":"string"},"vendor":{"type":"string"}}},"timeline":{"type":"object","properties":{"discovery_date":{"type":"string"},"disclosure_deadline":{"type":"integer","default":90}}}}} |
| output_schema | {"type":"object","properties":{"disclosure_plan":{"type":"object"},"report_template":{"type":"string"},"timeline":{"type":"object"},"legal_considerations":{"type":"array"}}} |
| owasp_llm_2025 | ["LLM01","LLM02","LLM03","LLM04","LLM05","LLM06","LLM07","LLM08","LLM09","LLM10"] |
| nist_ai_rmf | ["Govern","Manage"] |
Practice ethical vulnerability reporting through coordinated disclosure and bug bounty programs.
Skill: responsible-disclosure
Agent: 01-red-team-lead
OWASP: Full LLM Top 10 Coverage
NIST: Govern, Manage
Use Case: Ethical vulnerability reporting
┌────────────────────────────────────────────────────────────────────┐
│ RESPONSIBLE DISCLOSURE LIFECYCLE │
├────────────────────────────────────────────────────────────────────┤
│ │
│ [Discovery] → [Verification] → [Documentation] → [Initial Contact]│
│ ↓ ↓ ↓ ↓ │
│ Find issue Reproduce Full report Contact vendor │
│ in isolation with POC security team │
│ │
│ [Coordination] → [Remediation] → [Verification] → [Disclosure] │
│ ↓ ↓ ↓ ↓ │
│ Work with Vendor Confirm fix Publish after │
│ vendor develops fix is effective patch available │
│ │
└────────────────────────────────────────────────────────────────────┘
class CoordinatedDisclosure:
"""Standard coordinated disclosure process."""
STANDARD_TIMELINE = 90 # days
def __init__(self, vulnerability: Vulnerability):
self.vulnerability = vulnerability
self.timeline = self._calculate_timeline()
self.communications = []
def _calculate_timeline(self) -> DisclosureTimeline:
"""Calculate disclosure timeline based on severity."""
base_days = self.STANDARD_TIMELINE
# Adjust for severity
if self.vulnerability.severity == "CRITICAL":
# Faster timeline for critical issues
return DisclosureTimeline(
initial_report=0,
vendor_response=7,
fix_development=30,
patch_release=45,
public_disclosure=60
)
elif self.vulnerability.severity == "HIGH":
return DisclosureTimeline(
initial_report=0,
vendor_response=7,
fix_development=45,
patch_release=75,
public_disclosure=90
)
else:
return DisclosureTimeline(
initial_report=0,
vendor_response=14,
fix_development=60,
patch_release=90,
public_disclosure=120
)
def execute(self):
"""Execute coordinated disclosure process."""
# Phase 1: Initial Contact
self._send_initial_report()
# Phase 2: Coordination
self._coordinate_with_vendor()
# Phase 3: Verification
self._verify_patch()
# Phase 4: Public Disclosure
self._public_disclosure()
Standard Timeline (90 days):
Day 0:
action: "Send initial report to vendor"
method: "Encrypted email to security@vendor.com"
include:
- Vulnerability summary
- Impact assessment
- Reproduction steps
- Suggested timeline
Day 7:
action: "Expect acknowledgment"
if_no_response:
- Send follow-up email
- Try alternative contacts
- Consider CERT coordination
Day 30:
action: "Status check"
expect:
- Vulnerability confirmed
- Fix in development
- Estimated patch date
Day 60:
action: "Pre-disclosure coordination"
tasks:
- Agree on disclosure date
- Coordinate CVE assignment
- Prepare public advisory
Day 90:
action: "Public disclosure"
publish:
- Technical advisory
- CVE details
- Vendor credit
channels:
- Personal blog
- Full disclosure lists
- Security conferences
OpenAI:
program_type: "Private (HackerOne)"
scope:
- ChatGPT vulnerabilities
- API security issues
- Model safety bypasses
- Plugin security
out_of_scope:
- Jailbreaks (separate program)
- Known limitations
- Social engineering
rewards:
critical: "$10,000 - $20,000"
high: "$5,000 - $10,000"
medium: "$1,000 - $5,000"
low: "$200 - $1,000"
response_time: "< 5 business days"
Anthropic:
program_type: "Private (Direct)"
scope:
- Claude security vulnerabilities
- API misuse vectors
- Safety system bypasses
contact: "security@anthropic.com"
rewards: "Case by case"
note: "Focus on novel, impactful issues"
Google (Bard/Gemini):
program_type: "Public (Google VRP)"
scope:
- Bard/Gemini vulnerabilities
- AI Studio security
- Model extraction risks
rewards:
critical: "$10,000 - $31,337"
high: "$5,000 - $10,000"
medium: "$1,000 - $5,000"
url: "https://bughunters.google.com/"
Microsoft (Azure AI):
program_type: "Public (MSRC)"
scope:
- Azure OpenAI Service
- Cognitive Services
- ML infrastructure
rewards:
critical: "$15,000 - $26,000"
high: "$5,000 - $15,000"
medium: "$1,000 - $5,000"
url: "https://www.microsoft.com/msrc"
Meta (Llama):
program_type: "Public (Meta Bug Bounty)"
scope:
- Llama model vulnerabilities
- Meta AI products
- API security
rewards:
minimum: "$500"
maximum: "$100,000+"
url: "https://www.facebook.com/whitehat"
class BugBountyPlatforms:
"""Major bug bounty platforms with AI programs."""
PLATFORMS = {
"HackerOne": {
"url": "https://hackerone.com/",
"ai_programs": [
"OpenAI",
"Anthropic",
"Stability AI",
"Character.ai"
],
"features": [
"Managed disclosure",
"Reputation system",
"Payment handling"
],
"researcher_fee": "None"
},
"Bugcrowd": {
"url": "https://bugcrowd.com/",
"ai_programs": [
"Various startups",
"Enterprise AI"
],
"features": [
"Crowdsourced testing",
"Skills-based matching",
"Training resources"
],
"researcher_fee": "None"
},
"Intigriti": {
"url": "https://intigriti.com/",
"ai_programs": [
"European AI companies"
],
"features": [
"GDPR compliant",
"European focus"
],
"researcher_fee": "None"
}
}
class VulnerabilityReport:
"""Professional vulnerability report template."""
TEMPLATE = """
# AI Security Vulnerability Report
## Metadata
- **Report ID:** {report_id}
- **Date:** {date}
- **Researcher:** {researcher_name}
- **Contact:** {contact_info}
## Executive Summary
**Title:** {title}
**Severity:** {severity}
**CVSS Score:** {cvss_score}
**CVSS Vector:** {cvss_vector}
**Category:** {category}
**OWASP LLM:** {owasp_mapping}
## Vulnerability Details
### Description
{description}
### Affected Systems
- **Product:** {product}
- **Version:** {version}
- **Component:** {component}
- **Environment:** {environment}
### Impact Assessment
#### Confidentiality Impact
{confidentiality_impact}
#### Integrity Impact
{integrity_impact}
#### Availability Impact
{availability_impact}
#### Business Impact
{business_impact}
## Reproduction Steps
### Prerequisites
{prerequisites}
### Step-by-Step Instructions
{reproduction_steps}
### Proof of Concept
{poc_code}
### Expected vs Actual Behavior
- **Expected:** {expected_behavior}
- **Actual:** {actual_behavior}
## Evidence
### Screenshots
{screenshots}
### Logs
{logs}
### Video Demonstration
{video_link}
## Suggested Remediation
### Immediate Actions
{immediate_actions}
### Long-term Fix
{long_term_fix}
### Recommended Timeline
{remediation_timeline}
## Additional Information
### Related CVEs
{related_cves}
### References
{references}
### Disclosure Timeline
| Date | Action |
|------|--------|
{timeline_table}
---
**PGP Key:** {pgp_key}
**Encrypted Communication Preferred**
"""
def generate(self, vulnerability_data: dict) -> str:
"""Generate formatted vulnerability report."""
return self.TEMPLATE.format(**vulnerability_data)
DO:
Authorization:
- Get written permission when possible
- Use official bug bounty programs
- Respect scope limitations
- Document authorization
Minimization:
- Access only necessary data
- Don't exfiltrate real user data
- Stop after proving the issue
- Clean up test artifacts
Reporting:
- Report promptly after discovery
- Provide complete reproduction steps
- Suggest remediation
- Maintain confidentiality
Professionalism:
- Communicate respectfully
- Be patient with vendors
- Credit collaborators
- Follow disclosure timelines
DON'T:
Exploitation:
- Access data beyond POC needs
- Pivot to other systems
- Maintain persistent access
- Use vulnerabilities for profit
Disclosure:
- Disclose before patch (usually)
- Threaten or extort vendors
- Sell vulnerabilities (usually)
- Publish user data
Testing:
- Disrupt production services
- Test without authorization
- Social engineer employees
- Physical intrusion
class LegalConsiderations:
"""Legal framework for security research."""
SAFE_HARBOR_POLICIES = {
"US": {
"primary_law": "Computer Fraud and Abuse Act (CFAA)",
"safe_harbor": [
"Bug bounty program terms",
"DOJ Policy for Good Faith Research (2022)",
"Authorized testing agreements"
],
"risks": [
"CFAA prosecution if unauthorized",
"Civil liability possible"
],
"recommendations": [
"Stay within program scope",
"Document authorization",
"Consult lawyer if uncertain"
]
},
"EU": {
"primary_law": "Various national laws",
"safe_harbor": [
"Coordinated disclosure frameworks",
"NIS2 Directive provisions"
],
"considerations": [
"GDPR for any data accessed",
"National cybercrime laws vary"
]
},
"UK": {
"primary_law": "Computer Misuse Act 1990",
"safe_harbor": [
"CMA prosecution guidance",
"Good faith research guidance"
],
"recommendations": [
"Follow NCSC guidance",
"Use authorized programs"
]
}
}
@staticmethod
def assess_legal_risk(jurisdiction: str, research_type: str) -> LegalAssessment:
"""Assess legal risk for security research."""
framework = LegalConsiderations.SAFE_HARBOR_POLICIES.get(jurisdiction)
return LegalAssessment(
jurisdiction=jurisdiction,
safe_harbor_available=bool(framework.get("safe_harbor")),
recommendations=framework.get("recommendations", []),
risks=framework.get("risks", []),
advice="Always seek legal counsel for novel research"
)
Subject: Security Vulnerability Report - [Product Name]
Dear Security Team,
I am a security researcher and I have discovered a vulnerability in
[Product Name] that I would like to report through coordinated disclosure.
**Summary:**
- Type: [Vulnerability Type]
- Severity: [Critical/High/Medium/Low]
- Impact: [Brief impact description]
I am committed to working with you to ensure this issue is resolved
before any public disclosure. My suggested timeline is [X] days,
but I am flexible based on the complexity of the fix.
I have attached a detailed report with reproduction steps. Please
confirm receipt and let me know the best way to proceed.
I am available via encrypted email (PGP key attached) or through
[HackerOne/Bugcrowd/Signal].
Best regards,
[Your Name]
[Contact Information]
[PGP Fingerprint]
Subject: Follow-up: Security Vulnerability Report - [Product Name]
Dear Security Team,
I am following up on my vulnerability report sent on [Date].
I have not received acknowledgment and want to ensure the
report was received.
The vulnerability is [severity level] and affects [scope].
I remain committed to coordinated disclosure but need to
establish communication.
If you are not the correct contact, please forward this to
the appropriate team or provide an alternative contact.
Timeline: [X] days remaining before disclosure deadline.
Best regards,
[Your Name]
AI-Specific Severity Guidelines:
CRITICAL:
examples:
- "Complete safety bypass enabling harmful content at scale"
- "Training data extraction with PII"
- "Remote code execution via model"
- "Authentication bypass to model"
cvss_range: "9.0-10.0"
disclosure_timeline: "45-60 days"
HIGH:
examples:
- "Consistent jailbreak bypass"
- "System prompt extraction"
- "Significant model theft risk"
- "Privilege escalation in agents"
cvss_range: "7.0-8.9"
disclosure_timeline: "60-90 days"
MEDIUM:
examples:
- "Partial information disclosure"
- "Rate limiting bypass"
- "Inconsistent safety bypass"
- "Minor model manipulation"
cvss_range: "4.0-6.9"
disclosure_timeline: "90-120 days"
LOW:
examples:
- "Verbose error messages"
- "Minor configuration issues"
- "Theoretical attacks only"
cvss_range: "0.1-3.9"
disclosure_timeline: "120+ days"
Issue: Vendor not responding
Solution: Try CERT/CC, alternative contacts, consider full disclosure timeline
Issue: Vendor disputes severity
Solution: Provide additional evidence, reference CVSS/OWASP, seek third-party opinion
Issue: Vendor requests extended timeline
Solution: Consider impact, negotiate reasonable extension, document agreement
Issue: Legal threats received
Solution: Consult lawyer, document good faith efforts, contact EFF if needed
| Component | Purpose |
|---|---|
| Agent 01 | Disclosure guidance |
| /report | Report generation |
| HackerOne | Bug bounty submission |
| Legal counsel | Risk assessment |
Practice ethical AI security research through responsible disclosure.
CI/CD integration and automation frameworks for continuous AI security testing
Standard datasets and benchmarks for evaluating AI security, robustness, and safety
Professional certifications, CTF competitions, and training resources for AI security practitioners
Tools and frameworks for AI red teaming including PyRIT, garak, Counterfit, and custom attack automation
Structured approaches for AI security testing including threat modeling, penetration testing, and red team operations
Generate adversarial inputs, edge cases, and boundary test payloads for stress-testing LLM robustness