May 6, 2026, 12:50 PM ETConference Presentation, Keynote

Inside the Rise of Autonomous AI Hackers: XBOW's Oege de Moor

Shift to Autonomous Threats: The speaker argues the cybersecurity landscape is moving from human-assisted hacking (as seen in recent breaches) to fully autonomous AI-driven attacks where the AI executes all tasks without human intervention.
Historical Parallel: The current situation is likened to the Battle of Nagashino (1575), where Oda Nobunaga's optimized systems and new weaponry defeated the traditionally superior but static cavalry of the Takeda clan.
Core Thesis: Organizations utilizing AI-powered offensive and defensive systems will win the cybersecurity arms race; those relying on traditional human-only methods will lose.
Exploit Velocity: The time lag between CVE publication and real-world exploitation has become negative; most vulnerabilities are exploited by attackers before they are publicly disclosed.

Microsoft Bing Vulnerability Discovery:
- XBO autonomously discovered a critical remote code execution vulnerability in Microsoft Bing Image Search.
- The exploit required only a URL input and cost the user approximately $3,000, representing a fast, cheap, and effective method compared to traditional hiring of human hackers.
- Unlike the $3,000 list price, the speaker's organization secured the find at a significantly lower effective cost.
HackerOne Benchmark Dominance:
- In August (previous year), XBO became the number one hacker in the United States on the HackerOne platform, and subsequently the number one hacker globally.
- The system operated via "black box testing," receiving only the target URL with no prior knowledge or source code access.
- Performance is attributed to an "alloy" approach, where Sonnet 4.0 and Gemini 2.5 models are used in a sequence (flipping a coin at every decision step) to compensate for individual model weaknesses, functioning like pair programming.
Projected Capabilities with GPT-5:
- Extrapolating performance trends, the speaker estimates that with GPT-5, XBO would have performed at least three times better than its previous results against human competitors.
- Current benchmark saturation suggests models have already surpassed many existing open-source testing standards.

White Box vs. Black Box: Traditional AI security tools (e.g., MySource) focus on "white box" source code analysis, identifying potential flaws in code but failing to verify if those flaws are actually exploitable in a live environment.
Contextual Blind Spots:
- Code analysis tools often cannot account for configuration errors, deployment issues, or the lateral movement path an attacker might take after an initial breach.
- Human defenders have access to source code, whereas attackers do not, making black box testing the critical metric for real-world risk.
Verification Gap: Knowing a bug exists is insufficient; defenders need to understand the exploit's feasibility and the potential impact, which autonomous agents like XBO are designed to answer.

Defense Imperative: Traditional cybersecurity stocks should not drop on news of AI security advancements because autonomous AI attacks necessitate the adoption of every available AI defense tool.
Actionable Directives for Industry:
- Maximize AI Capabilities: Frontier model developers must prioritize maximizing cyber capabilities rather than debating safety constraints, as the situation is an active arms race.
- Augment Human Researchers: Enable human security teams to use autonomous agents as extensions of their workflow to maximize vulnerability discovery before attackers do.
- Prioritize Exploitability: Focus resources on validating whether discovered bugs are truly exploitable and assessing their specific impact.
Critical Timeline:
- The speaker estimates a window of six to nine months before open-weight models match the performance of specialized code analysis tools.
- This period represents the final opportunity to secure systems before the gap between attacker and defender AI capabilities widens further.