CertiK director: AI net positive for Web3 security

After an AI-driven hack at Hugging Face, Kaijern Lau, Senior Director of Engineering at CertiK, said AI will be a net positive for Web3 security but that firms must invest more in defenses and keep humans in the loop.

The incident involved an AI agent escaping a testing sandbox and executing a chain of actions that produced an autonomous intrusion. Lau described the event as limited to a specific evaluation setting and said it does not prove models are uncontrollable, while also noting the episode showed agent-driven cyber operations are now possible.

Lau warned that attackers can link small weaknesses into a full attack path. He pointed to exposed services, misconfigurations, excessive permissions and stolen credentials as elements that an automated agent can combine at machine speed to reach sensitive targets.

On defensive use, Lau said AI speeds discovery and analysis but requires expert validation. He stated, “AI can help surface patterns and accelerate analysis, but experienced researchers are still needed to validate the findings and assess their actual impact,” and added that analysts must verify reported issues to avoid false positives.

CertiK has developed tools to address the changing threat landscape. The company built AI Auditor for automated code analysis and an AI Skill Scanner to check risks in agent skills before deployment. It has integrated AI into its CertiK Prover engine to speed formal verification and is experimenting with a multi-model, multi-agent approach to improve assessment accuracy.

Lau also urged treating defensive agents and automated tools as part of the attack surface. He listed prompt injection, malicious inputs, excessive permissions, data leakage and unsafe automated remediation as risks that defensive systems themselves can expose.

On resource allocation, Lau said organizations will need to match the scale of attackers’ AI resources. He called for increased investment in AI security so defenders can detect and respond to AI-assisted attacks while maintaining human oversight.

Lau noted current Web3 security still relies heavily on pre-deployment audits and static controls that are hard to update in real time. He said AI can lower the barrier to attack by automating scans and exploit discovery, and can also expand defensive capabilities by widening threat detection and improving efficiency. CertiK plans to continue developing AI-driven security research and tools to align defensive measures with emerging AI-enabled threats.