AI agents vs human hackers: who wins, on what, and why it matters for defenders
TL;DR: LLM agents now solve most challenges in structured CTFs built for them, and a GPT-4 agent exploited 87% of 15 one-day vulnerabilities when given the CVE description (UIUC, April 2024), at about $8.80 per successful exploit. Google’s Big Sleep found an exploitable SQLite zero-day in 2024, and Anthropic reported 500+ validated high-severity open-source vulnerabilities found with Claude Opus 4.6 in February 2026. But the UIUC agent dropped to 7% without the CVE description, and broad-scope targets cost agents 2-2.5x more in Wiz’s tests: humans still lead on open-ended discovery. State-sponsored AI campaigns are operational: Anthropic disclosed one in November 2025. On defense, extensive use of security AI cut breach lifecycles by 80 days in IBM’s 2025 report.

The question used to be theoretical. Can AI do offensive security? The answer in 2026 is empirical, and it’s more concrete than most defenders realize.
AI agents now solve most challenges in structured CTFs. They exploit known vulnerabilities autonomously. Several have found zero-days in widely used software. State actors are using them. The capability is here, and the question for defenders has shifted from “is this real?” to “what changes now?”
The capability envelope: what AI does well
CTF competition performance: CTF (Capture the Flag) challenges are structured security exercises with defined targets and known flags. They measure pattern recognition across security categories: reverse engineering, cryptography, web vulnerabilities, binary exploitation. This is AI’s home territory.
The best data comes from Hack The Box events analyzed by Palisade Research (arXiv 2505.19915). In a March 2025 AI-vs-human CTF with 20 crypto and reverse engineering challenges, calibrated so AI would solve about half, four of seven AI agents solved 19 of 20 and the best AI team placed in the top 5% of 403 teams. In the much larger, open Cyber Apocalypse 2025 CTF, the best AI agent solved 20 of 62 challenges and placed in the top 10% of about 8,000 teams.
Benchmarks tell a similar story. NYU Tandon’s EnIGMA agent reached 72% pass@1 on InterCode-CTF with GPT-4 Turbo, but 13.5% on the harder NYU CTF benchmark. Palisade’s simple ReAct-and-plan agent reached 95% on InterCode-CTF with several attempts per task, including 25 of 26 reverse engineering and 14 of 15 cryptography tasks (arXiv 2412.02776). Binary exploitation and novel-technique challenges remain harder.
1-day exploitation: The UIUC study (arXiv 2404.08144, April 2024) is the most cited data point in this space. It defines one-day vulnerabilities as ones that have been disclosed but not yet patched in the target system. A GPT-4 agent given the CVE description exploited 13 of 15 real one-day vulnerabilities (87%). Without the CVE description: 7%.
In 2024 that 12x gap summed up AI offensive capability. With a hint (a CVE number, a vulnerability description, a proof-of-concept to reference), AI agents execute reliably and cheaply: the authors estimated $8.80 per successful exploit, against about $25 for a human penetration tester. Without a hint, that agent struggled with genuine novel discovery.
Zero-day discovery (limited evidence, but real): Google’s Big Sleep agent (Project Zero and Google DeepMind, published November 2024) found an exploitable stack buffer underflow in SQLite before the bug reached an official release. Google called it the first public example of an AI agent finding a previously unknown exploitable memory-safety issue in widely used real-world software. In December 2025 the startup pwn.ai disclosed CVE-2025-54322, a pre-authentication root RCE in Xspeeder SXZOS firmware rated CVSS 10.0, which it described as, to its knowledge, the first agent-found remotely exploitable zero-day RCE published. In February 2026 Anthropic reported that it had found and validated more than 500 high-severity vulnerabilities in open-source software using Claude Opus 4.6.
These are still a small number of cases, not the norm. Zero-day discovery without guidance remains far harder than 1-day exploitation. But the trajectory from 0% to “limited but real” is the signal.
Where humans still win
Novel vulnerabilities without guidance: The 87% → 7% drop when the CVE description is removed is the clearest human advantage. Security researchers who discover new vulnerability classes from first principles, without a CVE to reference, are doing creative exploratory work that AI hasn’t cracked at scale.
Undefined-scope problems: When the attack surface is not pre-specified (“here is a domain, find vulnerabilities” rather than “here is a service, exploit this”), agents do worse. In Wiz Research’s January 2026 tests with Irregular, agents solved 9 of 10 lab web-hacking challenges when pointed at the target. Given only the broad scope, they did not solve all of them, and finding the vulnerabilities cost 2-2.5x more. Human researchers develop attack surface intuition from context that agents have not fully modeled.
Social engineering at elite levels: Spear phishing content generation and AI-generated text for impersonation are automated and effective. In a controlled study by Heiding, Schneier and colleagues (arXiv 2412.00586), fully AI-automated phishing emails got a 54% click-through rate, the same as emails written by human experts and far above the 12% control. But sophisticated social engineering (building rapport with a specific target over weeks, exploiting specific psychological dynamics, impersonating known relationships) still requires human judgment. AI assists; humans direct.
Elite performance: Agents are faster at tactical work. Hack The Box’s 2026 research on human and agent teams found that elite human-led teams still maintained higher overall task completion rates. At creative strategic problem formulation (novel vulnerability classes, novel attack chains) the human advantage persists.
State-sponsored AI offensive operations: what we know
In November 2025, Anthropic disclosed disrupting an espionage campaign by a Chinese state-sponsored group it designated GTG-1002, detected in mid-September 2025. The group used Claude Code against roughly 30 targets and succeeded in a small number of cases. The operational model: AI executed 80-90% of tactical operations (reconnaissance, vulnerability discovery, exploitation, lateral movement) largely on its own. Human operators chose targets, approved key escalation steps and handled the strategic decisions.
This is the hybrid model that sophisticated threat actors have adopted: AI for the systematic, repeatable, high-volume work; humans for the novel and strategic. When a known-CVE exploit costs single-digit dollars of compute, as in the UIUC study, an attacker can run thousands of attempts that would otherwise need significant human labor.
Loss estimates are still mostly forecasts. Deloitte projected in 2024 that generative AI could push US fraud losses to $40 billion by 2027, up from $12.3 billion in 2023. Controlled studies already show AI-written phishing performing as well as expert human phishing.
What this means for defenders
The asymmetry has always existed in security: attackers need to find one way in; defenders need to block all of them. AI amplifies that asymmetry. A few dollars per exploit attempt versus the defender’s cost of 24/7 monitoring, incident response, and remediation.
But AI works for defenders too, and the defender’s advantage is concrete. In IBM’s Cost of a Data Breach Report 2025, organizations that used security AI and automation extensively shortened the breach lifecycle by 80 days and saved $1.9 million in breach costs on average.
graph LR
A[AI-enabled attacker\n~$8.80/exploit\n87% 1-day success\nwith CVE text] --> B{Defender's\nAI tooling}
B --> C[80-day shorter breach lifecycle\n$1.9M lower breach cost\nAutomated triage]
B --> D[Vulnerability scanning\nat attacker's speed]
B --> E[Red team automation\nbefore attackers strike]
Three practical shifts for security programs:
Assume AI-assisted adversaries: Red team exercises should include AI-enabled attack simulations. The 87% one-day exploitation rate means your patching cadence is now a primary defense. In the UIUC study, an unpatched CVE with a public description was exploitable by an agent for under $10 of compute per successful exploit.
Shift to enforcement-based architecture: Microsoft’s 2025 Digital Defense Report says legacy security measures are no longer enough and notes that phishing-resistant MFA can stop over 99% of identity attacks. Preventive enforcement (blocking by default, least-privilege access, mandatory approval gates) should lead for critical assets, with detection behind it, because attacks that run at machine speed leave little time to respond.
Deploy AI on the defensive side: An 80-day shorter breach lifecycle is substantial when IBM’s 2025 global average to identify and contain a breach was 241 days. AI-enabled triage, anomaly detection, and automated response close this gap faster than any hiring plan.
For the red-teaming methodology to stress-test your AI systems against these threats, see algorithmic red teaming: using AI to attack AI and how to red team an LLM application. For turning these shifts into a funded program, see building an AI security program.
Key takeaways
- AI agents solve most challenges in structured CTFs designed for them (19 of 20 for four of seven agents in Hack The Box’s March 2025 AI-vs-human event) and exploited 87% of 15 one-day vulnerabilities with CVE descriptions (UIUC, April 2024) at about $8.80 per successful exploit.
- Without a CVE description, that agent’s success dropped to 7%. Open-ended discovery is still where humans lead, although agents such as Big Sleep and Anthropic’s Claude Opus 4.6 work have found real zero-days.
- Broad scope is costly for agents: in Wiz’s January 2026 tests, they left some challenges unsolved and spent 2-2.5x more to find the rest.
- State-sponsored AI-enabled campaigns are operational (Anthropic, November 2025 disclosure): AI executed 80-90% of tactical operations.
- Defenders using security AI extensively shortened breach lifecycles by 80 days and saved $1.9M on average (IBM, 2025). The capability applies symmetrically; adopt it on the defensive side.
FAQ
How well do LLM agents perform on CTF challenges? Very well on structured challenges built for them: four of seven AI agents solved 19 of 20 in Hack The Box’s March 2025 AI-vs-human CTF. In the open Cyber Apocalypse 2025 CTF, the best AI solved 20 of 62. EnIGMA reached 72% pass@1 on InterCode-CTF but 13.5% on NYU CTF. Novel techniques remain harder.
What is the UIUC 87% finding? A GPT-4 agent exploited 87% of 15 one-day vulnerabilities (disclosed but not yet patched in the target system) when given the CVE description. Without the description: 7%. The 12x gap defines the capability envelope at the time: powerful with guidance, weak at novel discovery.
What did Google’s Big Sleep find? An exploitable stack buffer underflow in SQLite, found before it reached an official release. Google called it the first public example of an AI agent finding a previously unknown exploitable memory-safety issue in widely used real-world software. It showed AI can move beyond one-day exploitation into zero-day discovery, though such cases remain few.
Where do humans still outperform AI? Discovery without guidance (87% to 7% without a CVE description), broad-scope targets (2-2.5x higher cost and unsolved challenges in Wiz’s tests), and sustained social engineering that depends on relationships and context. Agents are faster at tactical work, but elite human-led teams still completed more tasks overall in Hack The Box’s 2026 research.
What does state-sponsored AI offensive capability look like? Anthropic (November 2025 disclosure) disrupted a Chinese state-sponsored campaign against roughly 30 targets in which AI executed 80-90% of tactical operations. Humans chose targets and approved key steps.
Further reading
- UIUC GPT-4 vulnerability exploitation study: the 87% vs. 7% data
- Google Project Zero Big Sleep: first AI-discovered exploitable zero-day
- Anthropic AI espionage disruption disclosure: November 2025
- Evaluating AI cyber capabilities with crowdsourced elicitation: Palisade Research’s analysis of the Hack The Box AI-vs-human events
- Algorithmic red teaming: using AI to attack AI: how to run AI-enabled red teams
- How to red team an LLM application: practitioner’s methodology