The AI Vulnerability Race Is Leaving Defenders Behind

AI is uncovering software vulnerabilities at a speed defenders cannot match. Zero-days can surface before security teams finish patching known flaws, creating a widening gap between discovery and repair.
That gap changes the entire security race. Frontier AI models now discover and exploit weaknesses that once demanded highly specialized researchers, while attackers can use the same capabilities to move through millions of lines of code before organisations finish assessing their vulnerability backlog.
AI Has Turned Vulnerability Hunting Into a Race
Vulnerability management followed a familiar rhythm for years: a flaw was discovered, someone reported it, a CVE was assigned, a patch became available, scanners flagged the issue, security teams assessed the risk, and engineers deployed the fix. Traditional scanners remain useful because they show which known vulnerabilities exist in an environment, but they cannot address weaknesses that have not been disclosed.
That model leaves defenders reacting to attackers’ timelines. Over the past four years, the average time from disclosure to exploitation has fallen from a month to less than a day, and nearly 900 vulnerabilities have been confirmed as exploited in 2026—nearly double the total for all of 2021.
AI is pushing that pressure earlier in the process. Agentic systems can search millions of lines of code with automated tools, identify complex chains of weaknesses, and examine access controls or business logic that traditional scanners may miss. Google has reported that attackers have already used agentic AI to inspect software updates, identify the security problems those updates address, and find ways to exploit them.
“I think vulnerability discovery without AI is basically dead,” said Steven Murdoch, a computer security researcher at University College London. He added that AI systems do not need to sleep and can work through many tasks in parallel, giving researchers and attackers a scale that human teams cannot match.
Daniel Gardham, a cybersecurity researcher at the University of Surrey, described the difference in scale: “They can set an AI agent off and come back in 20 minutes and have dozens of results already.” Capabilities that once required specialist expertise are becoming available to a much wider audience, including people with limited technical knowledge.
Finding the Flaw Is Easier Than Fixing It
The speed of discovery does not guarantee a safe repair. Over half of AI-generated patches failed to fix the original flaw or introduced new ones, while AI-generated code is frequently insecure and can take longer to repair because developers may not understand how it was constructed.
AI-assisted developers produce code three times as fast, but that output can expand the security workload. “The vulnerability backlog has always existed. The challenge now is that it’s a whole lot bigger,” said Sandeep Johri, CEO of Checkmarx. “With LLMs being weaponised, you can’t afford that. You have to work through your backlog and keep it as small as possible.”
PHP2Shell shows what this new race looks like in practice. Research from Wiz, Calif and OpenAI’s Codex Security team found that frontier AI helped a single researcher uncover critical flaws in the PHP engine in about a week. PHP supports applications built on WordPress, Drupal and Joomla, and the runtime underpins three-quarters of the world’s websites.
The two critical vulnerabilities, CVE-2026-109618 and CVE-2026-109619, were rated CVSS 9.2. They involved memory-safety flaws in PHP itself, and researchers demonstrated remote code execution and memory disclosure against widely used applications, including WordPress, Drupal and Joomla.
The flaws had been known for years, yet exploiting them once required analysis at the edge of what human researchers could achieve. AI delivered that analysis at a speed that would have been difficult to imagine a few months ago, turning PHP2Shell into a clear warning for every team that relies only on disclosed CVEs.
More Findings Still Demand Better Decisions
AI does not produce one shared map of risk. Echo tested Anthropic’s Claude Mythos 5.1, OpenAI’s GPT-5.6 Sol through Daybreak, and Z.ai’s GLM-5.3 against three open-source projects using the same vulnerability-hunting framework and validation process. Of 121 distinct potential issues, only one was found by all three models, while 83% were found by just one.
Zhipu AI’s open-weight GLM-5.3 developed working end-to-end exploits in 50 out of 410 attempts on the ExploitBench benchmark. Claude Mythos Preview managed 56. Those results show that models can uncover serious weaknesses, but they also show why one model’s answer cannot serve as the final security decision.
AI can raise false alarms, and even the same model can produce very different analyses of the same code. A list containing real problems, uncertain findings and false positives is not security by itself; it is a larger workload. Faster scanning matters, but accurate scanning with fewer false positives matters more.
The response must also cover more than software created by in-house developers. Legacy platforms such as PHP, vendor SaaS products and cloud-native applications all require examination, and teams must verify fixes from end to end across build, deployment and runtime behaviour.
Only around 20 per cent of companies have formal AI governance policies, leaving businesses exposed to the risks of shadow AI use. The future of application security will involve AI, but AI alone cannot provide certainty. A hybrid human-machine model combines deterministic analysis from skilled people with AI’s probabilistic reasoning, context awareness and advanced analysis.
“The life for the defenders who are trying to fix things has become significantly harder,” Murdoch said. The winning security teams will not simply find more flaws; they will prove which findings matter, verify that repairs work, and keep pace before attackers turn discovery into exploitation.
Based on
- AI is Finding Vulnerabilities at Warp Speed. Fixing Them is a Whole Other Story. — unite.ai
- Three Leading AI Models Hunted the Same Code for Vulnerabilities and Agreed on Just 1 in 121 | Markets Insider — markets.businessinsider.com
- The AI security risk gap: building best practice for AppSec | The Independent — independent.co.uk
- AI is turbocharging cyberattacks — can computer scientists catch up? | Nature — nature.com
- After Black October, OffSec Turns PHPH2Shell and Other New Vulnerabilities Into Hands-On Labs | Currency News | Financial and Business News | Markets Insider — markets.businessinsider.com




