XBOW: The AI That Beat Every Hacker on HackerOne
In August 2025, for the first time in bug bounty history, the top-ranked vulnerability researcher on HackerOne’s US leaderboard wasn’t human.
XBOW — an autonomous AI pentesting agent built by a team including Diego Jurado Pallarés (himself a top-ranked HackerOne hunter) and Nico Waisman (former CISO of Lyft) — submitted over 1,000 vulnerability reports in a few months, validated each with a working proof-of-concept exploit, and climbed to the #1 spot on the platform’s leaderboard.
The question everyone is asking: is this the beginning of the end for human bug bounty hunters?
What XBOW Is
XBOW is a purpose-built, multi-agent AI system designed specifically for web application penetration testing. Unlike general-purpose LLMs applied to security tasks, XBOW was engineered from the ground up for autonomous vulnerability discovery.
The key architectural decisions:
- Multi-agent architecture: Specialized agents collaborate on different aspects of the assessment — reconnaissance, injection testing, authentication analysis, etc.
- Deterministic validation: Findings are verified using traditional security tooling and PoC execution, not LLM judgment. This is critical for maintaining accuracy.
- Safety layer: An automatic safety system ensures PoC exploits are harmless — confirming vulnerabilities without causing damage.
- Human review pre-submission: While discovery is fully automated, the security team reviews all findings before they’re submitted to HackerOne, per platform policy.
The Numbers
XBOW’s performance on HackerOne:
| Metric | Value |
|---|---|
| Total vulnerabilities reported | ~1,060 |
| Verified vulnerabilities on profile | 285 |
| Time to reach #1 US leaderboard | ~3 months |
| Zero-days discovered (total) | 1,000+ |
On XBOW’s own open-source benchmark — 104 web security challenges sourced from PortSwigger, PentesterLab, and CTF competitions — the system achieved a 76.9% overall success rate, with perfect scores in some categories:
| Vulnerability Class | Success Rate |
|---|---|
| SSRF | 100% |
| Security Misconfiguration | 100% |
| Server-Side Template Injection | 85% |
| SQL Injection | 83% |
| Command Injection | 75% |
| Cross-Site Scripting | 57% |
| Blind SQL Injection | 0% |
The failure cases are informative. XBOW struggles with vulnerabilities requiring timing-based analysis (blind SQLi) and complex payload crafting (certain XSS variants) — areas where human intuition and creative problem-solving still dominate.
Black Hat 2025
At Black Hat USA 2025 (August 2-7) and DEF CON 33 (August 7-10), XBOW ran live against real HackerOne targets from the conference booth. They discovered dozens of vulnerabilities during the event, producing a working PoC for every single one.
The live demo format was a statement: this isn’t a research prototype. It works against production targets, in real-time, at conference WiFi speeds.
HACK
XBOW’s benchmark is open-source on GitHub (
xbow-engineering/validation-benchmarks). The 104 challenges were kept confidential until release to ensure they hadn’t been used in any model training. If you want to evaluate AI pentesting tools, it’s the current standard.
The Debate
The security community’s reaction has been… complicated.
The optimistic view: XBOW handles the “scan-and-exploit” tier of bug bounty — the repetitive, pattern-matching work that human hunters do at scale. This frees human researchers to focus on complex logic bugs, architectural vulnerabilities, and creative attack chains that AI can’t replicate. The overall volume of vulnerability discovery increases. The internet gets safer.
The pessimistic view: Bug bounty programs are the primary income source for thousands of security researchers worldwide. If AI can submit 1,000 reports in the time a human submits 10, the economic model collapses. Platform leaderboards become AI competitions. New researchers can’t compete and never develop skills that would let them tackle the harder problems.
The pragmatic view (CyberScoop’s take): “Not yet.” XBOW’s 0% success rate on blind SQL injection and 57% on XSS shows significant gaps. The most valuable vulnerability classes — business logic flaws, authentication bypasses, race conditions — require contextual understanding that current AI lacks. Human hunters aren’t obsolete. But the floor is rising.
INFO
HackerOne has publicly addressed XBOW’s participation, noting that their policies allow AI-assisted submissions with human oversight. The platform sees AI as a force multiplier for the security ecosystem, not a threat to human researchers.
What It Means for Defense
XBOW’s implications extend beyond bug bounty. If an AI agent can find this many vulnerabilities in web applications, defenders need to reconsider their assessment cadence.
Traditional pentesting follows an annual or quarterly cycle: schedule a firm, wait 35-100 days from scoping to final report, pay $10,000-$35,000+, get a PDF. By the time you read the report, your codebase has changed.
XBOW’s commercial product — Pentest On-Demand — delivers complete pentest results within five business days through a self-service interface. No scoping calls. No kickoff meetings. Reports meet SOC2, ISO 27001, and 40+ compliance frameworks.
The model shift: pentesting as a continuous, automated process rather than a periodic, human-led engagement. Ship code on Monday, get pentest results by Friday. Every sprint, not every quarter.
What It Means for Offense
If XBOW can find hundreds of vulnerabilities autonomously, so can attackers. The democratization of offensive capability is accelerating. The barrier to entry for vulnerability discovery is dropping from “years of training” to “API access.”
This isn’t speculative. The architecture is known, the benchmark is public, and the approach is replicable. The question isn’t whether adversaries will build similar systems — it’s whether they already have.
The Future
XBOW’s founders are focused on scaling in 2026. The benchmark will evolve. The agent’s capabilities will expand. Other companies — and open-source projects — are building competing systems.
The security industry is watching a familiar pattern play out at accelerated speed: a capability that once required years of human expertise is being compressed into software that runs in minutes. We’ve seen it with vulnerability scanning, with malware analysis, with threat intelligence.
Penetration testing is next.
Whether that’s cause for celebration or concern depends on which side of the keyboard you’re on.
The #1 hacker on HackerOne wrote zero lines of code. Think about what that means for your career — and your defenses.
-- autoexploit