lonkero is a Rust-built web security scanner with 125+ modules, ML-driven false-positive reduction, and an AI agent mode, aimed at penetration testers working on authorized targets.
bountyyfi/lonkero — professional-grade, modular web security scanner written in Rust | |
| Category | web application vulnerability scanner / automated pentesting tool |
| Primary Use | Authorized penetration tests and web app assessments: XSS, SQLi, SSRF, JWT, race conditions, WebSocket issues, and supply-chain checks via lonkero scan |
| Safe Use | Use only against systems you own or have written authorization to test: internal labs, engagement scopes, bug bounty programs, and CI/CD security validation |
| Telemetry Note | The scanner generates distinctive request patterns (canary-prefixed probes, multi-value SLEEP timing tests, quote-oscillation sequences) that WAF logs and anomaly detection can flag; request-volume heuristics make it observable to defenders |
bountyyfi/lonkero describes itself as a production-grade web security scanner for real penetration testing, written in Rust on 1.85+ and currently at version 3.7.0. The repository sits at roughly 1,100 stars with a proprietary license, which matters operationally: this is not a permissively licensed tool you can freely embed into a commercial product, but it is publicly documented and built for professional use. The README is unusually detailed and reads like a changelog-driven design document, with each major version introducing a coherent detection philosophy rather than an accumulation of modules.
The core pitch is context-aware scanning under the banner of Intelligent Mode. Instead of spraying thousands of payloads at every parameter, lonkero performs tech detection, endpoint deduplication, and per-parameter risk scoring before deciding what to test. Parameters named like password, cmd, or query get escalated to heavier payload sets, while page, limit, and sort get minimal treatment. The README claims a 5% false-positive rate against an industry baseline of 20-30%, and notably states that when tech detection fails, the fallback layer runs more tests, not fewer — a sensible design choice since fingerprinting errors are a classic source of scanner blind spots.
The XSS engine received a full rewrite in v3.6, replacing Chrome-based detection with what the authors call a mathematical proof-based approach. The scanner sends a baseline request with a CANARY value, locates reflection points, then sends a single probe with escaping metacharacters to characterize filtering behavior. Sixteen reflection contexts are recognized — HTML body, JS strings, attributes, event handlers, javascript: URLs, comments, CSS — and exploitability is derived from context plus escape analysis rather than browser observation. This drops the per-parameter cost to 2-3 requests, which the README claims is roughly 300x faster than browser-based scanning while still achieving 99% accuracy on reflected HTML and attribute XSS.
Payload volume is user-controllable through the --payload-intensity flag, ranging from minimal at 50 payloads per parameter up to maximum at 12,450+, with auto as the risk-scored default. This is a thoughtful knob for authorized workflows: minimal maps naturally to CI/CD pipelines where scan time and request budgets are constrained, while extended and maximum suit deep-dive engagements against dedicated staging environments. The filter heuristics are also deliberately conservative, skipping CSRF tokens, pagination parameters, and boolean flags where reflection is structurally impossible.
The v3.2 OOBZero engine is the most technically interesting section. Traditional blind SQL injection detection depends on out-of-band infrastructure — Collaborator, Interactsh, custom DNS — which is operational overhead and, on restricted engagements, may be prohibited. lonkero instead calibrates timing with multiple SLEEP values and computes Pearson correlation, treating r > 0.95 as confirmation that the database is responding to injected commands. It then layers boolean data extraction, quote-oscillation pattern matching, and Bayesian weighting across timing, content length, and boolean differentials to converge on certainty without any external callback. The trade-off, honestly acknowledged in the README, is more requests per finding.
v3.1 broadened coverage in ways that reflect real-world scanner maturity. The skip logic for static versus dynamic targets was corrected so that Cloudflare Workers, Vercel Functions, and Netlify Functions are properly tested. SSRF testing was extended to POST JSON and form-encoded bodies rather than only query strings. New scanners include second-order injection (payload stored on one endpoint, execution detected on another), an auth flow tester covering session fixation, password-reset IDOR, and MFA bypass patterns, plus expanded JWT weak-secret checks and WebSocket CSWSH testing with nine origin bypass techniques.
One scanner stands out as genuinely novel: the README Invisible Prompt Injection detector, built on the vendor's own research at bountyyfi/invisible-prompt-injection. It looks for instructions hidden in README.md files via HTML comments, [//]: # (...) reference links, and hidden reference definitions — content invisible when rendered but fully readable by LLM coding assistants processing raw markdown. Findings are classified under CWE-1059 as a supply-chain issue, with heuristic flags for dependency injection patterns like require(, URL hijacking, and prompt-override phrases. For teams adopting AI-assisted development, this doubles as a defensive supply-chain audit capability, and the remediation guidance (strip comments before LLM processing, audit raw source) is sound.
Version 3.7 adds lonkero ai, a conversational agent layer where an LLM drives the 94+ scanner modules through natural language. Rather than firing every module at once, the agent reasons like a pentester: recon first, identify interesting targets, then run targeted scans with specific modules and chain findings. The README is explicit that AI keys are bring-your-own, meaning no vendor-hosted inference is required and engagement data does not transit a third party's API unless you choose it. Operationally, this positions lonkero closer to agent-driven testing frameworks than to legacy crawl-and-blast scanners.
From a telemetry and detection standpoint, lonkero is loud in recognizable ways. Canary-prefixed parameter values, multi-value SLEEP timing sequences, quote-oscillation request patterns, and single-packet raw TCP/TLS timing probes all leave distinctive signatures in WAF and application logs. Defenders running lab replicas of the tool can fingerprint its behavior and build correlation rules around it; conversely, operators on authorized engagements should expect request-rate anomalies and coordinate scan windows with the client's monitoring team.
Modern stack coverage is a recurring theme: Next.js, React, GraphQL, gRPC, WebSocket, and HTTP/3 are all listed as supported, and endpoint discovery ships with 244+ patterns for API, admin, debug, and tool paths. The repository publishes a passing test badge with 95% coverage, a release workflow, and a browser extension branded EXTENSIO, suggesting the project invests in CI discipline rather than shipping a research prototype. For authorized professionals evaluating a scanner that treats false positives as a first-class engineering problem, lonkero is worth a structured trial in a lab before deployment on client scope.
bountyyfi/lonkero.Educational analysis for authorized security professionals. Use only in controlled, authorized environments.
Related coverage
0 comentários:
Post a Comment
Note: Only a member of this blog may post a comment.