CertiK’s Intel3D report, published on October 5, 2026, argues that agentic AI is transitioning from a support role into an operational workforce across crypto security, compliance, and financial-crime investigations. These autonomous systems can reason through multiple steps, invoke external tools and APIs, gather evidence, act in live environments, and evaluate results with limited human input. As a consequence, agentic agents are taking on smart contract analysis, live transaction monitoring, cross-chain fund tracing, and transactional risk screening that previously required human analysts.
The report frames this shift in part as a response to attack velocity: flash-loan exploits and rapid laundering via mixers, bridges and OTC conversions compress response windows into seconds or hours. CertiK cites the Bybit case—where 86.29% of the stolen ETH was converted to Bitcoin within a month—to illustrate how quickly funds can move. The wider loss context in CertiK’s data includes $768.4 million lost in September across 97 incidents and roughly $2.68 billion in total losses through September 2026 under the firm’s methodology, while H1 2026 losses were reported at $1.32 billion. The report also recalls regulatory pressure, noting that AML fines exceeded $900 million in H1 2025 according to earlier CertiK analysis.
Agentic systems are described as capable of more than detection: in mature setups, detection can trigger automated responses within the same block window, such as pausing vulnerable contract functions, activating circuit breakers, or freezing compromised keys. Agents can continuously follow stolen assets across chains and update address clustering as new activity appears, combining cross-chain data into single investigations rather than analyzing networks in isolation. The report also notes that autonomous agents are entering compliance workflows, preparing regulatory reports, reconciling onchain and offchain records, and performing real-time address and transaction screening.
CertiK emphasizes that autonomy brings new risks. Agents can produce confident but incorrect outputs, human reviewers may grow complacent, and attackers can exploit AI-driven tools or manipulate agents via prompt injection. Security agents themselves become targets if they have authority to interact with sensitive systems. To manage these risks the firm recommends treating agents as workforce participants with defined scopes of authority, named human owners, complete audit trails of inputs, reasoning and actions, escalation procedures, capped autonomy, adversarial testing, and rigorous behavioral monitoring. The report stresses that responsibility remains with organizations and the humans who configure and supervise agents, and warns that without these controls automation may replace familiar problems with failures that are harder to explain.