Blockchain security firm CertiK says autonomous AI agents can now monitor on-chain transactions, trace stolen crypto across chains, and automatically trigger defensive responses—capabilities the firm argues are necessary to combat attacks that move faster than human analysts can respond.
In its Intel3D report published October 5, 2026, CertiK outlined how agentic AI software can operate continuously without human prompting. According to the report, these agents can spot threats and pause vulnerable contracts within the same block window, prepare regulatory reports, and activate circuit breakers—the crypto equivalent of trading halts during market crashes.
Why Speed Matters
CertiK frames rapid automated response as essential against flash loan exploits, which borrow large sums, manipulate a protocol, and repay the loan within a single transaction. By the time humans notice, the firm notes, the funds are typically gone.
The report also highlighted the challenge of tracing laundering routes. Stolen funds often move through mixers that blend coins to obscure their origin and bridges that transfer assets between blockchains. CertiK said manual tracing struggles to keep pace with these complex paths.
The firm cited the Bybit exploit as an example of the window problem: 86.29% of stolen ETH from that attack was converted to Bitcoin within a month. CertiK also referenced major breaches at Bitget and Liquid Network in September 2026 to emphasize how short the response window typically is after a hack.
Accountability Remains with Humans
CertiK was explicit that AI agents should operate within defined roles and authority levels, with organizations assigning a human owner to each agent who remains accountable for its decisions.
The report identified specific failure modes: an agent could freeze a legitimate transaction or file an inaccurate report, either of which carries real consequences including customer dissatisfaction and regulatory problems.
To mitigate risk, CertiK recommended detailed audit trails, rigorous testing protocols, and capped autonomy to address both errors and adversarial attacks where bad actors attempt to manipulate the AI systems.
