Ask PhantomYerra (AI Pentester)
Ask PhantomYerra is your AI penetration tester. It plans the engagement, selects from 87+ engines across 16 attack surfaces, interprets results, adapts its strategy in real time, chains vulnerabilities into attack paths, runs zero-day detection, and writes the final report — all autonomously.
How Ask PhantomYerra Drives the Pentest
Traditional "AI-assisted" tools add AI as a layer on top of existing scan output — generating summaries or suggesting fixes after the fact. PhantomYerra is architecturally different: Ask PhantomYerra is the orchestrator. Every engine PhantomYerra runs is a callable function that the AI invokes based on its own reasoning about what to test next. With 87+ engines across 16 surfaces — including an 11-engine Zero-Day Detection Suite — it has broader coverage than any human tester working alone.
Ask PhantomYerra's Role as Your AI Pentester
Ask PhantomYerra receives the engagement brief (target, scope, engagement type, organisation context) and then calls tools as functions — exactly as a senior penetration tester would direct a team. It decides which engines to run, in what order, and adapts its plan based on each result. When it finds something interesting, it goes deeper. When a path is a dead end, it pivots. When source code is in scope, it activates the Zero-Day Detection Suite to find vulnerabilities that have never been assigned a CVE.
// Simplified: how the AI orchestrates PhantomYerra
tools = [run_port_scan, run_vuln_scan, run_injection_test, run_directory_scan,
read_finding, add_finding, run_sast, check_idor, ...]
ai_engine.run(
"You are conducting an authorised pentest of [TARGET].
Plan and execute systematically. Chain findings into
attack paths. Prove every vulnerability with evidence."
)
// AI then calls tools, reads results, calls more tools...
The 5-Phase AI Pipeline
Pre-Recon - Target Intelligence
The AI engine validates scope, gathers passive intelligence without touching the target: WHOIS, DNS records, certificate transparency logs, ASN data, known breaches (HaveIBeenPwned), public CVEs for identified technologies, and threat intelligence feeds. It builds an initial attack hypothesis before any active probe is sent.
Active Recon - Surface Mapping
The AI engine directs active reconnaissance: port scanning, service fingerprinting, web crawling, API endpoint discovery, technology stack identification. It prioritises based on Pre-Recon findings, if it found a CVE for the web server version, it tests that first.
Vulnerability Analysis - Targeted Testing
With a full attack surface map, the AI engine selects the highest-impact attack vectors and tests them in priority order. It generates context-aware payloads tailored to the specific technology stack: not generic signature-based tests. SQLi payloads for PostgreSQL differ from MySQL; XSS payloads adapt to the CSP policy detected.
Exploitation - Proof of Impact
For each confirmed vulnerability, the AI engine executes the full exploitation chain to prove maximum impact. It does not stop at "SQL injection detected": it extracts data, escalates privileges, pivots to adjacent systems where permitted by scope, and documents every step. Evidence is captured automatically at each stage.
Reporting: AI-Written Executive and Technical Report
The AI engine writes the complete pentest report: executive summary with business impact, full technical findings with evidence, attack chain narrative, risk-prioritised remediation roadmap, and compliance mapping (OWASP Top 10, PCI DSS, ISO 27001, NIST). All evidence is embedded. The report reads as if written by a senior consultant - because it was.
Zero-Day Detection Suite
When source code or mobile APKs are in scope, Ask PhantomYerra automatically activates the Zero-Day Detection Suite — 11 dedicated engines that find vulnerabilities invisible to pattern-based and CVE-signature scanners. These engines detect logic-level flaws: vulnerabilities that exist in your specific codebase but have never been assigned a CVE and never appear in any signature database.
SAST Zero-Day Engines (7)
| Engine | What It Finds | CWEs / Severity |
|---|---|---|
| Interprocedural Taint Flow | Cross-file BFS taint propagation — tracks untrusted data from 20+ source types to 25+ sink types across function call boundaries | CWE-89/78/79/22/94/502/601 |
| Race Condition Detection | TOCTOU, broken double-checked locking, mutex misuse, unsafe temp file patterns — AST-level analysis across concurrent access paths | CWE-362/367/833/820/377 |
| Crypto Oracle Analysis | Padding oracle (CBC + distinguishable exceptions), timing oracle (non-constant-time HMAC), ECB mode, GCM nonce reuse, weak KDF — 5 languages | CWE-327/385/208/330/916 |
| Auth Chain Analysis | JWT alg:none (CVSS 9.8), RS256→HS256 downgrade, IDOR patterns, MFA bypass flows across multi-step auth chains | CWE-287/384/639/345 · CVSS up to 9.8 |
| Deserialization Gadget Finder | 5 languages, gadget chain detection, automatic ysoserial/phpggc PoC generation for Java/PHP targets | CWE-502 · CVSS 9.8 |
| Supply Chain Analysis | Levenshtein typosquatting (50+ packages, ≤2 distance), known malicious package database, postinstall script analysis | CWE-1104/1357 |
| AI Adversarial Scanning | 5 AI passes: business logic, parser differential, trust boundary, state machine, type confusion — routes through 7-provider chain with graceful degradation | Business logic · Multi-surface |
Mobile Zero-Day Engines (4)
| Engine | What It Finds | CWEs / Severity |
|---|---|---|
| DEX Bytecode Analysis | Dynamic class loading, SSL pinning bypass, AES/ECB use, struct-level DEX string table parsing — Smali + Java analysis | CWE-295/470/327/925 |
| Intent Fuzzer | Static AndroidManifest exported component detection + ADB dynamic fuzzing, ContentProvider SQL injection, path traversal | CWE-926/89/22/20 |
| WebView Bridge Analyzer | addJavascriptInterface API<17 (CVSS 9.8), setAllowUniversalAccessFromFileURLs sandbox escape (CVSS 8.8), Intent extras injection | CWE-749/346/73/601 · CVSS up to 9.8 |
| IPC Violation Detector | Binder/AIDL interface exposure, ContentProvider access without permission, PendingIntent escalation, PreferenceActivity fragment injection | CWE-862/89/22/284/926/927 |
Non-fatal by design: All 11 zero-day engines are wrapped individually — a single engine encountering an unsupported language or parsing edge case never aborts the parent scan. The remaining engines continue and all confirmed findings are included in the final report.
The 7-Provider AI Chain
Ask PhantomYerra routes AI requests through a 7-provider chain in priority order, with automatic failover. A scan never fails due to a single provider being unavailable or rate-limited.
| Priority | Provider | Best For |
|---|---|---|
| 1 — Primary | Anthropic Claude | Complex reasoning, zero-day AI passes, report narrative writing |
| 2 | OpenAI | Code analysis, payload generation |
| 3 | Google Gemini | Large-context analysis, multi-file SAST |
| 4 | Groq | Ultra-low-latency inference, high-throughput scans |
| 5 | Together AI | Open-source model access, cost-efficient bulk analysis |
| 6 | Ollama | Local on-device, air-gapped deployments |
| 7 | LM Studio | Local on-device, Windows/macOS GUI model management |
Cloud Mode vs Local Mode
Ask PhantomYerra works in both cloud AI and local on-device AI modes. Choose based on your data policy and capability requirements.
Cloud Mode: 7-Provider AI Chain
- Anthropic → OpenAI → Google → Groq → Together (automatic failover)
- Highest reasoning capability for complex zero-day analysis
- Best for report narrative writing and AI adversarial passes
- PrivacyFilter anonymizes all target data before sending
- Requires internet connection
- Small per-scan cost (varies by depth)
- Best for: standard commercial engagements
Local Mode - On-Device AI
- Zero data transmitted - 100% local
- Works fully air-gapped
- Local models approach cloud AI quality
- No per-scan cost after model download
- Slower on CPU; fast on GPU
- Best for: classified, healthcare, finance engagements
- See: Local-Only Mode guide
Configuring your AI Key
Key from License Server (Automatic)
If your PhantomYerra license tier includes an AI API key, it is downloaded and configured automatically when you activate your license. No manual setup required. The key is shown (masked) on the License page under AI Configuration.
Manual Key Entry
If you are using your own AI API key: open Settings → AI Configuration and paste your API key. PhantomYerra validates the key with a minimal test call before saving. The key is encrypted with AES-256-GCM at rest.
Credit Monitoring
PhantomYerra monitors AI API credit consumption. When credits are running low, a non-blocking toast notification appears. When credits are exhausted, a blocking screen offers the option to add a new key or switch to local on-device AI mode.
Setting Up Local On-Device AI for Air-Gapped Mode
PhantomYerra supports a fully local AI mode for air-gapped deployments using Ollama (providers 6 and 7 in the chain). To enable local AI:
- Install Ollama or LM Studio on your machine (both supported natively in the 7-provider chain).
- Pull a recommended model appropriate for your hardware — high-VRAM GPU recommended for best performance. Recommended:
deepseek-r1:70b(complex reasoning) orcodellama:34b(code analysis). - In PhantomYerra: Settings → AI Configuration → Local-Only Mode → Enable — PhantomYerra auto-detects the local runtime and selects the best available model.
After this setup, disconnect from the internet — PhantomYerra continues to operate with full AI capability using the on-device model exclusively. The Zero-Day Detection Suite's AI Adversarial Scanning engine also routes through the local provider in air-gapped mode.
Reading the AI Decision Log
In Automated AI mode, the AI Brain Log panel in the scan dashboard shows Ask PhantomYerra's real-time reasoning. This is a live window into why the AI is making each decision — invaluable for understanding the engagement and learning from Ask PhantomYerra's approach.
[CALL] run_vuln_scan(template="CVE-2022-32224", target="[TARGET_URL_1]")
[RESULT] Template matched - server responded to YAML probe, possible deserialization endpoint at /api/import
[THINK] Match is promising but not confirmed. Need to verify with actual deserialization payload and OOB callback to prove RCE.
[CALL] run_deserialization_exploit(endpoint="/api/import", oob_host="oob.phantomyerra.io")
[RESULT] OOB callback received: DNS lookup from target IP confirms code execution. RCE CONFIRMED.
[THINK] RCE confirmed. Now determining impact: what can we access? Check for /etc/passwd, environment variables, cloud metadata endpoint.
[CALL] extract_data(command="cat /etc/passwd && env | grep -i key")
[RESULT] Extracted: /etc/passwd (32 users), AWS_ACCESS_KEY_ID present in environment
[ADAPT] Critical finding: AWS key in environment. Pivoting to cloud escalation. Checking IAM permissions...
Manual Override - Guiding the AI
Even in Automated AI mode, you retain full control. You can intervene at any point without stopping the scan.
| Action | How | When to Use |
|---|---|---|
| Pause AI and take manual control | Click "Pause AI" in the scan dashboard | When you want to manually test a specific endpoint the AI hasn't reached yet |
| Direct AI to a specific surface | Type a command in the AI chat panel: "Focus on the GraphQL API at /graphql" | When you know something is likely vulnerable and want AI to prioritise it |
| Approve before exploit | Enable "Confirm before exploitation" in scan settings | For sensitive targets where you want human approval before destructive actions |
| Override tool selection | Right-click any AI tool call in the log → "Run manually instead" | When you want to adjust parameters or use a different payload set |
| Resume AI from current point | Click "Resume AI": it picks up from where you paused | After manual testing, hand back control to AI to continue systematic coverage |
Best practice: Let Ask PhantomYerra run Phase 1–3 fully automatically (recon + vulnerability analysis + zero-day detection). Manually review findings at Phase 4 (exploitation) for high-severity issues. Resume AI for Phase 5 (reporting). This gives you maximum coverage with human oversight on the highest-impact actions.