Akira is a complete offensive security skill suite that runs natively inside your AI coding environment.
No Docker. No server. No 40-tool pre-install hell. No hallucinated findings.
Six phases. One chain. One deliverable report.
/plan-engagement -> /recon -> /secrets -> /exploit -> /zerodayhunt -> /triage -> /report
Each phase reads structured output from the previous phase and writes for the next. Every finding requires direct HTTP evidence. No proof = no finding. Always.
PentestGPT has 12k stars and a documented hallucination problem. HexStrike has tool coverage but no phase handoffs. Trail of Bits has CI/CD coverage but no engagement lifecycle.
Akira has all of it - and none of their weaknesses.
| Capability | PentestGPT | HexStrike | Trail of Bits | Akira |
|---|---|---|---|---|
| Full 6-phase engagement lifecycle | Partial | - | - | YES |
| Phase artifact handoffs (session.json) | - | - | - | YES |
| Anti-hallucination evidence gate | - | - | - | YES |
| Confidence scoring per finding (0-100) | - | - | - | YES |
| False positive verification gate | - | - | Partial | YES |
| Active Directory full chain (BloodHound -> DCSync) | - | Partial | - | YES |
| OAuth/OIDC exploitation suite | - | - | - | YES |
| Race conditions (single-packet attack) | - | - | - | YES |
| Cloud audit (AWS + GCP + Azure + K8s) | - | Partial | - | YES |
| CI/CD GitHub Actions attack vectors | - | - | YES | Coming Month 2 |
| CTF mode (HackTheBox, TryHackMe) | YES | - | - | YES |
| Native in Claude Code, Gemini, Cursor | Partial | - | - | YES |
| One-command tool bootstrap | - | - | - | YES |
| Real bug bounty proof (updated weekly) | - | - | - | YES |
| Free + MIT | YES | Partial | YES | YES |
Option 1 - Clone and install skills (recommended):
git clone https://github.com/Kalp1774/akira
cd akira && bash install.shOption 2 - Install tools too (nuclei, dalfox, subfinder, etc.):
git clone https://github.com/Kalp1774/akira
cd akira && bash install.sh && bash bootstrap.shThat's it. Open Claude Code (or your AI environment), type /plan-engagement target.com, and go.
Verify install:
/plan-engagement example.com
You should see the engagement plan and session.json initialized.
| Platform | Install | Skill Syntax |
|---|---|---|
| Claude Code | install.sh copies to ~/.claude/skills/ |
/plan-engagement, /recon, etc. |
| Gemini CLI | Add platform-adapters/GEMINI.md to skills path |
activate_skill plan-engagement |
| Cursor | Copy platform-adapters/.cursor/rules/akira.mdc |
Cursor rules auto-activate |
| Codex (OpenAI) | See platform-adapters/.codex/INSTALL.md |
Reference via AGENTS.md |
| GitHub Copilot CLI | AGENTS.md pattern | Natural language trigger |
| Any agent | AGENTS.md in repo root | Plain text skill invocation |
| Skill | Phase | What It Does |
|---|---|---|
/plan-engagement |
0 | Scope definition, PTT generation, session.json init |
/recon |
1 | Subdomains, live hosts, ports, URLs, tech stack fingerprint |
/secrets |
2 | API keys, tokens, credentials in JS/source/git/Postman |
/exploit |
3 | XSS (dalfox), SQLi (sqlmap), nuclei scan, deserialization, SSTI, XXE, NoSQLi |
/zerodayhunt |
3+ | Chain attacks, JWT confusion, SSRF->IAM, WAF bypass, type juggling |
/triage |
4 | Severity clustering, confidence scoring, FP verification, deduplication |
/report |
5 | Pentest report or HackerOne/Bugcrowd submission format |
| Skill | What It Does |
|---|---|
/ad-attacks |
BloodHound path analysis, Kerberoasting, AS-REP, DCSync, Golden/Silver Ticket, ADCS ESC1-8 |
/oauth-attacks |
Redirect URI bypass, CSRF on OAuth, PKCE downgrade, JWT confusion, implicit flow token theft |
/race-conditions |
HTTP/2 single-packet attack, coupon reuse, double-spend, OTP bypass, TOCTOU |
/cloud-audit |
AWS SSRF->IAM, S3 enum, IAM privesc, GCP service account, Azure RBAC, K8s API |
/ctf |
HackTheBox/TryHackMe, web/crypto/pwn/RE/forensics/OSINT/stego methodology |
Every skill reads structured output from the previous phase. No intelligence is lost between phases.
plan-engagement
writes -> session.json (target, scope, tech stack, attack priority tree)
recon
reads -> session.json
writes -> interesting_recon.md (live hosts, ports, URLs, headers)
secrets
reads -> interesting_recon.md
writes -> interesting_secrets.md (verified API keys, tokens, credentials)
exploit
reads -> interesting_recon.md + interesting_secrets.md
writes -> interesting_exploit.md (confirmed vulns with HTTP evidence)
zerodayhunt
reads -> ALL prior outputs
writes -> interesting_zerodayhunt.md (chains, zero-days, critical paths)
triage
reads -> ALL interesting_*.md files
writes -> triage.md (severity-clustered, confidence-scored, FP-verified)
report
reads -> triage.md + session.json
writes -> report-YYYY-MM-DD.md (final deliverable)
Akira's biggest technical differentiator. Every finding must pass the evidence gate before it reaches the report.
Evidence Requirements (enforced in every skill):
- Every CONFIRMED finding must quote the exact HTTP response body proving it
- Empty 200 response body = WAF catch-all = NOT a finding
{"status":403}in a 200 response = WAF block = NOT a finding- A header existing does not mean the header is exploitable
Confidence Scoring (0-100):
| Score | Meaning |
|---|---|
| 90-100 | Reproducible, full data exfil or proven RCE |
| 70-89 | Strong evidence, not fully chained yet |
| 50-69 | Behavioral indicator, no data proof |
| < 50 | Speculative - auto-excluded from report |
Findings below 70 are automatically downgraded to POTENTIAL and excluded from final report.
Real anonymized findings made with Akira. Updated weekly.
See FINDINGS.md for full writeups.
| # | Type | Severity | Platform | Bounty | Skill Chain |
|---|---|---|---|---|---|
| 1 | SSRF -> AWS IAM Credential Extraction | Critical | HackerOne | $2,500 | /recon -> /exploit -> /cloud-audit |
| 2 | OAuth Open Redirect -> Authorization Code Interception | Critical | Bugcrowd | $1,800 | /recon -> /oauth-attacks |
| 3 | Race Condition: Coupon Applied 7x Simultaneously | High | Private | $800 | /race-conditions |
| 4 | Strapi SSRF Bypass + MIME Fail-Open (CVE filed) | Critical | CVE | - | /zerodayhunt |
| 5 | JWT RS256->HS256 Algorithm Confusion -> Admin Access | Critical | HackerOne | $1,500 | /zerodayhunt |
Akira ships new skills every month. Here's what's coming:
Month 1 - SHIPPED (v1.0.0)
- Core 7-phase lifecycle
- ad-attacks (BloodHound -> DCSync full chain)
- oauth-attacks (open redirect -> ATO chains)
- race-conditions (HTTP/2 single-packet)
- cloud-audit (AWS/GCP/Azure/K8s)
- ctf (HTB/THM/PicoCTF)
- Phase handoff system (session.json)
- Anti-hallucination evidence gate
Month 2 - Coming
-
graphql- Introspection abuse, batching, field-level authz bypass -
deserialization- Java/PHP/Python/.NET gadget chains (ysoserial, PHPGGC) -
prototype-pollution- Node.js client + server RCE chains -
supply-chain- Dependency confusion, namespace squatting, typosquatting -
ci-cd-audit- 9 GitHub Actions attack vectors (Trail of Bits methodology + more)
Month 3 - Coming
- Akira Context Engine - auto-pulls CVEs for detected stack + HackerOne disclosed reports
-
cache-attacks- Cache poisoning + cache deception unified playbook -
csp-bypass- CSP escape chains, JSONP, nonce reuse
Month 4 - Coming
-
mobile- Android APK analysis, Firebase misconfig, iOS IPA, Frida certificate pinning bypass -
burp-integration- Native Burp Suite MCP integration (Repeater, Intruder, proxy history)
Month 6 - Coming
- Akira Brain - Persistent attack tree across sessions, cross-engagement memory
-
postmap-recon- PostmapDB integration (leaked Postman collections -> live secrets) -
red-team- MITRE ATT&CK full emulation
These are the four infrastructure upgrades that turn Akira from a great skill suite into an unbeatable autonomous offensive engine. Not new skills — foundational rewrites.
No more "command not found." No more broken local environments.
Instead of asking users to install 50 tools, Akira ships as a single container:
docker pull ghcr.io/kalp1774/akira-engine
docker run -it akira-engine /plan-engagement target.comEvery tool (nuclei, nmap, sqlmap, subfinder, dalfox, httpx, katana, trufflehog) is pre-installed, path-configured, and version-locked inside. The AI knows exactly what it has. The user installs nothing.
Why it matters: The current install + bootstrap model breaks on Windows, WSL, and hardened macOS. A container eliminates the entire class of "my environment is broken" failures. It also makes CI/CD security testing trivial — just pull and run.
Sending 5,000 lines of raw Nmap output to a cloud LLM is engineering malpractice.
A lightweight Python parser runs locally before any output reaches the AI. It reads raw tool output and extracts only the signal:
Raw nmap output: 5,000 lines → Parser → 12 lines sent to AI
Open 80 (Apache 2.4.49)
Open 443 (nginx/1.18)
CVE-2021-41773 detected
Why it matters: API cost drops by ~90% per engagement. The AI's context window stays clean — no 4,990 lines of closed ports and timing noise. Smarter decisions, smaller bills. This is the difference between a $1 scan and a $10 scan at scale.
The AI shouldn't just run commands — it should feel the target's defenses and adapt.
A real-time monitor sits between the terminal and the AI. If it detects 3 consecutive 403 Forbidden or 429 Too Many Requests responses, it auto-pauses the current phase and injects a decision:
[STEALTH GOVERNOR] 3x 403 detected on api.target.com
Switching to residential proxy. Adding 5s delay. Resuming with --rate-limit 10.
Why it matters: Right now Akira runs hot until it gets banned. The Governor turns it into a patient hunter — it learns the target's WAF fingerprint in real-time, adjusts aggression, and never burns its own IP. This is what separates a script kiddie tool from professional tradecraft.
The context crash is the biggest unsolved problem in AI-assisted pentesting.
Every finding — IPs, API keys, headers, versions, credentials, behavioral observations — gets indexed into a local vector database (ChromaDB) as it's discovered. Phases query the vault instead of relying on in-context memory:
# Exploit phase queries its own vault instead of reading files
results = vault.query("any SSRF endpoints found during recon?")
results = vault.query("AWS credentials from secrets phase")Why it matters: A pentest can run for 3 days. Right now, after 2 hours the context window fills and the AI "forgets" the first recon findings. With the Memory Vault, every detail from Day 1 is queryable on Day 3. The "Context Crash" disappears entirely.
The current phase chain is a line. Real attacks are graphs.
When the AI discovers a new attack surface mid-engagement — a new IP, an internal subdomain, an exposed service — it should be able to fork a background task instead of abandoning the current thread:
Main thread: /exploit api.target.com (continues)
Fork spawned: /recon internal.target.com (runs in background)
Fork spawned: /secrets new-js-bundle.js (runs in background)
Forks report back when complete. The main thread picks up their findings automatically.
Why it matters: Right now finding something new mid-phase means either ignoring it or starting over. The Graph Engine turns Akira from a strict checklist into a collaborative team — multiple threads working in parallel, the way a real red team operates.
Akira is free and MIT licensed. If it helped you find a bug or win a CTF, consider supporting development:
- GitHub Sponsors: @Kalp1774
- Buy Me a Coffee: buymeacoffee.com/kalpmodi
- Ko-fi: ko-fi.com/kalpmodi
Every sponsorship goes directly into new skills, technique research, and real bug bounty hunting to keep FINDINGS.md alive.
Found a technique that should be in Akira? Bug in a skill? New tool that belongs here?
- Fork the repo
- Add or improve a skill in
skills/<skill-name>/SKILL.md - Open a PR with a brief explanation of what it covers
- If it's from a real finding - include a link to the public disclosure (anonymized is fine)
Attribution added to FINDINGS.md for all confirmed-finding contributions.
See SECURITY.md for how to report skill bugs, suggest techniques, or submit findings made with Akira.
Submit Akira to these lists after launch to drive organic traffic:
- awesome-security - open a PR to the Tools section
- awesome-pentest - open a PR
- awesome-claude-code - open a PR
- awesome-hacking - open a PR
- awesome-bug-bounty - open a PR
- HackerOne Resources - submit tool
- kali.org/tools - longer term goal
Akira is for authorized security testing only.
Use only on:
- Bug bounty programs you are enrolled in
- Systems you own or have explicit written permission to test
- CTF competitions
Unauthorized testing is illegal. The authors are not responsible for misuse.
Built for bug hunters, by bug hunters.
Star this repo to stay updated when new skills ship.