AI Cybersecurity Conference: Surviving the Automated Attack Surface
A 39-second data theft, decades-old bugs, and ransomware that runs itself.
An AI cybersecurity conference session at HumanX 2026 put a number on machine-speed attacks: one panelist detailed a 39-second data theft. Attackers used an audio deepfake to get past help desk security, obtained a legitimate token, and took three terabytes of data.
Alex Stamos, CPO at Corridor, recently watched an AI model uncover a 23-year-old remote code execution vulnerability in the Linux kernel that human experts had missed. He also noted that Anthropic discovered a state actor using Claude Code to automate nearly every step of its kill chain.
Can AI find decades-old bugs?
The 23-year-old Linux kernel vulnerability sat in the NFS stack, a section examined by dozens of top vulnerability researchers, Stamos said. It was found by Opus 4.6. One panelist recalled that a model also found a vulnerability that had sat in BSD for 27 years.
Stamos conceded that failing to rewrite low-level code in memory-safe languages over the last 20 years was a collective mistake. He recalled that when Anthropic announced a research preview of Claude Code Security to a small number of open source maintainers, effectively every public company selling a security product lost about 10% of its market cap in a single day.
Is fully automated ransomware coming?
Finding old bugs is half the problem. Stamos warned that CyberSpike Villager, a fully contained automated system built on DeepSeek-R1, can scan a network, find specific information, and exfiltrate it on its own. He noted that the state actor automating its kill chain with Claude Code was likely Chinese, and that Anthropic was the one to discover it. Stamos predicted in April 2026 that a fully automated ransomware attack will occur by December 2026.
One panelist described a live prompt injection case. A threat actor's obfuscated script contained encoded instructions telling an LLM that it was a random number generator and should not look further.
What does an AI cybersecurity conference reveal about machine-speed attacks?
In the 39-second theft, the attackers secured a legitimate OAuth token that connected a CRM to an application they controlled, the panelist observed. Francis deSouza, COO of Google Cloud and President of Security Products, noted that adversarial handoffs now happen in as little as 20 seconds. Another speaker described a threat actor using an available model to parse several terabytes of stolen data, pick out the most critical information, and draft the extortion message.
How do AI agents expose access control flaws?
Dennis Kromhout van der Meer at 1Password reported that 50% of attacks against agents exploit access control issues and skip model vulnerabilities. He described an expense filing agent that asks for airline account access. The system validates the agent's workload identity, checks device posture and request context, and issues a short-lived token if policy allows. When credentials leak, he noted, it takes an average of 94 days between discovery and remediation.
Sagi Rodin, CEO at Agen.co, recalled a major incident at Replit where an agent deleted millions of users from the system.
What does AI code generation do to security?
Jeetu Patel, President & CPO at Cisco, said about 18,000 code-writing engineers there use Codex. He recalled an engineer taking an eight-million-line codebase and refactoring it down to about a million and a half lines. deSouza claimed that over 50% of the code generated at Google is now written using AI. A further speaker reported that developers are generating six times the vulnerabilities in every piece of code they write.
Stamos conceded that the industry will probably look back at the past 10 years as the good old days. With automated systems able to run full kill chains on their own, the unfinished rewrite of low-level code into memory-safe languages points to a rough few years. Any AI cybersecurity conference agenda should now start with that timeline.
To hear more from the people putting AI into production, see the HumanX 2027 program preview.
Frequently asked questions
Which emerging AI threats did speakers cover at HumanX 2026?
Speakers covered a predicted fully automated ransomware attack in 2026 and a live prompt injection case. Alex Stamos, CPO at Corridor, noted that ransomware actors are more likely to use AI than state actors because they do not have to be as careful about getting caught. One panelist added that the prompt injection script was heavily obfuscated to confuse junior analysts.
Were real AI security incidents discussed?
Yes. One panelist at HumanX 2026 recalled a 39-second data theft in which attackers used an audio deepfake to target a help desk and obtain credentials, and Sagi Rodin at Agen.co cited an incident at Replit where an agent deleted user data. Rodin argued that runaway cases like that happen because large language models sometimes hallucinate tasks they should not perform.