A security assessment by blockchain cybersecurity firm Hacken has raised a practical warning for the fast-growing market for AI trading agents: blocking withdrawals alone would not prevent a compromised or poorly controlled bot from causing severe losses.
Hacken’s analysis, included in a recent report on AI-assisted cryptocurrency trading, said an agent that can place orders may still be manipulated into buying tokens at inflated prices, entering oversized positions, or accumulating exposure that exceeds a user’s intended risk. In each case, funds can be lost through trading activity without ever leaving the account.
The finding places attention on the permissions granted to AI systems rather than on withdrawal settings alone. An agent does not need authority to transfer assets externally to damage an account. Access to spot or derivatives markets can be enough if the system is able to place large orders, use leverage, or repeatedly trade into thin liquidity.
Trading permissions create risks beyond withdrawals
AI trading tools are increasingly used to summarize market news, analyze charts, identify potential opportunities, and generate trading ideas. The step from research assistance to autonomous execution introduces a different set of risks because software begins acting on live balances rather than offering information for a human to assess.
Hacken said safeguards should be placed outside the agent itself. That approach would prevent a trading model from changing its own limits or bypassing restrictions when it receives flawed, manipulated, or malicious instructions.
The firm identified caps on individual orders, total positions, and account permissions as core controls. Such restrictions could limit the damage from a single bad trade or a sequence of automated trades made during a volatile market move.
A cap on order size, for example, would stop an agent from deploying an unexpectedly large share of available funds in one transaction. A total-position limit would restrict how much exposure the account can build across several orders. Permission controls could prevent the agent from accessing certain assets, markets, or high-risk features such as leveraged positions.
These measures matter in crypto markets where prices can move rapidly and liquidity may vary sharply between assets. A bot can execute faster than its user can intervene, particularly if it is allowed to monitor markets and place orders continuously.
External controls are designed to survive agent failures
Hacken also described security testing intended to block unwanted actions before execution. The goal is to treat an AI agent as a component with limited authority rather than as a system that should be trusted to police itself.
That distinction becomes especially relevant when an agent receives information from external sources. A model that reads market news, social-media posts, token data, or other online material could encounter misleading signals or malicious instructions designed to influence its behavior. The report’s security framework focuses on ensuring that the final trading action remains subject to fixed controls.
In practice, an effective design would separate an agent’s ability to recommend or request a trade from the system that authorizes it. The agent could identify a potential order, but a rules-based layer would check whether the trade exceeds pre-set capital, position, leverage, or loss thresholds before it reaches the market.
This structure would also give users a clearer way to audit what their automated tools are doing. Users can review which actions were proposed, which were blocked, and whether a limit was reached. That record can help identify a faulty strategy, an unexpected market condition, or behavior that does not match the user’s instructions.
Users appear more comfortable with AI research than autonomous trading
The report’s survey findings showed a clear divide between the use of AI as a research tool and willingness to hand over direct execution. Respondents reported using AI more often for tasks such as summarizing news, analyzing markets, and finding potential trading opportunities than for placing orders automatically.
That pattern reflects the practical difference between using AI to process information and allowing it to control capital. Market research can be checked, ignored, or compared with other sources. An automated order can be filled in seconds, potentially before a trader notices that the assumptions behind it have changed.
The report also found that protective controls affected respondents’ willingness to consider automated trading. Interest increased when scenarios included notifications, the ability to pause activity, permission revocation, and automatic stops tied to loss limits. Yet respondents generally remained cautious about the share of account assets they would delegate to an automated system.
That hesitation is consistent with Hacken’s warning. Automation may reduce the time required to monitor markets or act on a trading strategy, but it also compresses the time available to catch errors. Limits on capital and permissions therefore serve as a circuit breaker when a model behaves unpredictably or reacts to poor information.
Exchange design will shape how much control users delegate
The report also found differing preferences on whether safeguards should be optional settings selected by users or minimum requirements applied by an exchange. Many respondents favored having control over the safety settings they enable, while support also remained strong for mandatory measures such as risk assessments, independent audits, and non-disableable limits on capital or losses.
That creates a design challenge for trading platforms. Highly restrictive rules may frustrate experienced users seeking flexibility, while fully optional settings can leave less experienced traders exposed to tools they do not fully understand.
A workable model would combine baseline protections with adjustable risk settings. Platforms could require clear limits on delegated authority while allowing users to choose tighter caps on funds, assets, order size, leverage, and permitted strategies.
Hacken’s assessment suggests that the most effective protections are likely to be the least glamorous: hard limits, isolated permissions, pre-trade checks, and a reliable way to stop the system. As AI trading agents move from market analysis toward execution, the ability to constrain a bot’s actions may become more valuable than the sophistication of its market predictions.
Concerned about AI trading risks? Learn how security controls help safeguard automated strategies and funds.
Disclaimer: The content on this page is provided for general informational purposes only and does not represent the views or financial advice of Toobit. We make no guarantees regarding the accuracy or completeness of this information and shall not be held liable for any errors, omissions, or outcomes resulting from its use. Investing in digital assets involves risk; users should independently evaluate their financial situation and the risks involved. For further details, please consult our Terms of Service and Risk Disclosure.
