代理式AI正成为加密安全领域的一部分：Certik

Proxy AI is taking over manual work in cybersecurity, anti-money laundering and compliance

Proxy AI is gradually taking over work that was previously handled directly by human analysts, covering areas such as cybersecurity, anti-money laundering and compliance. Currently, autonomous systems are able to investigate threats, track funds, and respond to security incidents with little human intervention.

From assistive tools to autonomous operators

Proxy-based artificial intelligence has gone beyond the mere marking stage of abnormal activity. According to CertiK's latest report, autonomous systems are now able to conduct multi-step reasoning, invoke external tools and APIs, collect evidence, and take action in a real-time environment, then evaluate the results to decide the next step. This emerging model is described as an "AI safety workforce" in which autonomous systems operate within established roles, while human experts turn to supervision, quality control and accountability.

In the past decade, AI for cybersecurity and compliance has mainly played a supporting role. Machine learning systems identify anomalies, natural language tools help summarize alerts or prepare reports, which are ultimately reviewed and acted upon by humans. However, proxy AI changes the proportion of this process that does not require direct human participation. In the Security Operations Center, autonomous systems can investigate abnormal logins, obtain device fingerprints and location history, compare them with threat intelligence, decide whether to suspend accounts and perform the suspension operation. Its reasoning process will be recorded for human review.

报告指出，此类系统应被视为具有明确职责的劳动力参与者，而非传统的软件工具。尽管责任仍归属于部署这些系统的组织以及配置和监督它们的人类，但组织若将自主AI视为普通软件工具，可能会因缺乏足够的审计轨迹、行为监控或升级控制而面临风险。

应对加密攻击速度的自动化需求

推动安全工作自动化的压力部分源于加密攻击的速度。闪电贷漏洞可在数秒内抽干协议资金，被盗资产也可在数小时内通过多个地址、桥接服务和混合服务转移。尽管头部损失有所下降，但实际损失依然巨大。例如，2026年上半年加密货币损失达13.2亿美元，其中第二季度钱包成为最大的攻击手段。

智能合约安全的新范式

在Web3安全领域，自主系统正在承担以往需要经验丰富的工程师才能完成的任务。传统的智能合约审计结合手动代码审查、静态分析、符号执行和模糊测试工具，但仍需人工判断潜在漏洞是否为真实安全问题。如今，代理式系统可以遍历合约调用图，分析跨合约的状态变化和外部调用，查找重入、预言机操纵、访问控制和不安全升级机制等漏洞。

In addition, AI is also applied to formal verification. Agents can generate formal specifications and test them against contract behavior, reducing the manual workload of formal method engineers. Still, human auditors are part of the process, mainly responsible for verifying AI-generated discoveries, investigating new economic or game theory attack methods, and checking for blind spots in the automated security system itself.

Real-time transaction monitoring and cross-chain fund tracking

CertiK's report extends automated security beyond the audit phase and into real-time transaction monitoring. AI systems can monitor pending and confirmed blockchain transactions and detect utilization patterns such as lightning loan attacks, oracle manipulation, and abnormal liquidity extraction. In more advanced settings, detection can trigger automatic responses within the same block window, such as suspending a vulnerable contract function, activating a circuit breaker, or freezing a corrupted administrator key.

In terms of fund tracking, proxy systems can continuously track the path of asset movement, rather than just rebuilding the path after a transaction occurs. By analyzing transaction times, overlapping counterparties, and Gas fee behavior, AI can more accurately determine whether an address cluster is controlled by the same entity. For cross-chain money laundering, the agency system can integrate activities on different chains into a single investigation and continuously track the flow of funds. For example, money laundering cases related to the Bybit vulnerability showed that stolen ETH was converted into Bitcoin in a short period of time through mixers, bridges, and over-the-counter brokers.

Automation challenges in compliance

Autonomous systems are also responsible for compliance functions. CertiK said that "Know Your Address"(KYA) and "Know Your Transactions"(KYT) screenings can now be completed by AI agents before blockchain transactions are settled. Agents assess the risks of addresses or transactions in real time, allowing platforms to identify potential illicit asset exposures before completing the transfer.

Regulatory reporting can also be automated. Proxy systems can extract blockchain information, reconcile it with off-chain records, and prepare reports involving obligations such as travel rule data sharing and stablecoin reserve certificates. As countries move from formulating encryption frameworks to actively enforcing laws, anti-money laundering fines have increased significantly, placing a huge regulatory burden on encryption companies.

New risks and responsibilities brought by autonomy

Giving AI systems the right to act introduces new security risks. Proxy systems may express a high degree of confidence while generating erroneous messages, leading to erroneous reports of suspicious activity or misjudging the effectiveness of formal specifications. If human reviewers increasingly trust automated output, it may become more difficult to catch such errors. At the same time, attackers are also using AI to accelerate vulnerability discovery, automated reconnaissance and create more convincing social engineering activities.

Security agents themselves can also be targeted because they often have permissions to interact with sensitive systems. Prompt injection or input manipulation may lead to discretionary approval of fraudulent transactions or the disabling of legitimate security controls. The report emphasizes that when autonomous actions cause damage, legal liability remains not fully resolved in various jurisdictions. The organization deploying the system and the people responsible for configuration supervision still have responsibilities.

CertiK recommends maintaining a complete audit trail of agent inputs, reasoning, and actions, setting clear independent decision-making limits, countering manipulation tests, and appointing one person for each agent accountable to the human owner. As MetaMask and others introduce wallets that allow autonomous agents to execute transactions, ensuring transparency and controllability of agent behavior has become crucial.