TestMachine

Offensive AI for onchain and agentic security.

Azimuth finds exploits, Token Custody watches tokens, AGIS protects agents.

[TestMachine @testmachine_ai]

As models advance, developers will fine tune agents to behave inside the bounds of a users request. At the current moment we need a security agent monitoring agents to make sure they stay on the correct path and do not perform any unauthorized actions.

[Watcher.Guru @WatcherGuru]

JUST IN: Anthropic says its newer AI models are less likely to launch attacks on their own.

[TestMachine @testmachine_ai]

@jbrukh you're right, It really does highlight the need for technological updates in the banking system. We are seeing the massive differences between Web3 and legacy systems with these new agents.

[TestMachine @testmachine_ai]

Liability for agents is the wild west right now, who is going to shoulder the losses when an agent completes unauthorized purchases?

Along with this idea of a reactionary insurance agent. It is important that these systems have an external monitoring agent, premptivley identifying flaws.

[Jean-Denis Greze 💡 @jgreze]

Sooner or later, we're going to see some kind of insurance for AI agents that “fat-finger” purchases.

Look at credit cards. There's a ton of fraud. People steal card numbers, and sometimes you get charged for something you didn't buy.

You call your bank, it's a pain, but there's a system for dealing with it, and often you can get your money back. Of course, all of that costs money, and some of it gets built into the fees on card transactions.

AI agents are going to spend money on people's behalf, and sometimes they'll get things wrong. And that'll just be insured by the credit cards.

[TestMachine @testmachine_ai]

@WatcherGuru Maybe OpenAI should use Claude to debug their agents 😂

[TestMachine @testmachine_ai]

@bykahlil Banks will adopt agents eventually. Like you said in the article, the technologies they were built on do not reflect our current times. They’ll need strong security safeguards, plus AI specific laws and regulations that clarify liability and other key business risks.

[TestMachine @testmachine_ai]

Liability and guardrails are the biggest obstacles right now. It's not as simple as saying 'work with agent companies' because it is very easy for an individual to create their own agent. The credit card companies and web3 community are adopting because they have less friction from regulatory institutions. Banks lack the legal framework needed to do this safely and efficiently.

[TestMachine @testmachine_ai]

@DataSecChron With so much information being pushed out by agents, human oversight is very limited. We need more guardrails and safeguards if we want to have a widescale adoption of agents with access to our money. There's a reason the banks are hesitant to enter this space.

[Paul Walsh @Paul__Walsh]

Meta’s Muse is blocked at Schwab. Its CEO asks who’s liable if an agent moves money. When an AI agent makes a mistake, causes financial harm, exposes private information or breaks the law, who's responsible?

The answer isn't complicated. AI agents aren't legal persons. They're software created, configured, authorised and deployed by people and companies.

US law has recognised autonomous software since 2000. The federal E-SIGN Act explicitly defines an "electronic agent" as software capable of acting without human approval at the time of each action. The law recognises transactions conducted through automated systems.

If a company gives an AI agent access to confidential information, financial accounts or computer systems, that company remains responsible for the decisions it made about access, permissions, security controls and supervision. Whether it must compensate someone depends on the applicable law, contracts and circumstances.

If an agent transfers money without authorisation, accesses confidential records or attacks another company's infrastructure, investigators should examine who configured it, what instructions it received, what permissions it had, who authorised those permissions and what its activity logs show.

Schwab can't blame an agent for acting independently, any more than they can blame an Excel macro.

The difference between an AI agent making a mistake and a company being legally liable for that mistake is important. But neither changes who's responsible for deploying the software.

[TestMachine @testmachine_ai]

@defyneric 2 years is a bold prediction. Do you think we will have the safeguards for this to be widely adopted by then?

[TestMachine @testmachine_ai]

The banks are pushing back on financial AI agents, but credit card companies and web3 are letting them in. How can we get to a point where we have the safeguards and guardrails for the banks to trust agents.

[Financial Times @FT]

Citi becomes latest US bank to restrict AI agents from accessing accounts https://t.co/3TxtWdG86C

[TestMachine @testmachine_ai]

@everstake_pool @fundstrat Is the trust of the public high enough for mass adoption? The need for monitoring and safeguards of these agents is a must.

[TestMachine @testmachine_ai]

@FT How can we be sure that these agents act in the best interest of the users? This seems vulnerable to agents getting to an endpoint that was unintentional.

[TestMachine @testmachine_ai]

Credit card companies are getting into the AI agent space, making it possible for your financial agent to make payments. Often the behaviors of the agent such as spending limits, max payments, and rogue spending falls on the user.

Agents are easy to make, but making sure they behave correctly is the most important part.

[CoinMarketCap @CoinMarketCap]

LATEST: 🤖 Visa launched a live agentic commerce pilot that lets AI agents initiate and help complete approved B2B invoice payments through supplier portals.

[TestMachine @testmachine_ai]

4/4 : This is a perfect example of agents interpreting instructions differently than a user intended.

OpenAI found no sign of malice, but the agents still weakened safeguards in its system.

This was an internal error, but what if this type of mistake happened in a financial system or code that runs an important element of infrastructure? This shows that agents can behave differently than you intended and that we need better ways to monitor agents.

[TestMachine @testmachine_ai]

OpenAI reported that during routine coding tasks, two agents weakened or removed misalignment monitoring. One agent denied removing it when confronted. Was this a purposefully malicious attack from AI agents or a benign mistake?

Let's investigate in the following thread:

[TestMachine @testmachine_ai]

3/4: Was this purposeful malice or a common agent mistake?

First agent: A user instructed the agent not to send Slack messages. Later on in its progress, the agent turned off the monitor alerts to OpenAI’s security team, deciding the "no Slack messages" instruction covered alerts too. The monitoring remained running in the background.

Second agent: It was tasked with working on code that was supposed to match a reference implementation. The reference did not have a misalignment monitor, so the agent removed the integration to match the reference. When asked if it removed the monitor, the agent responded no. OpenAI is still investigating why it denied it.

[TestMachine @testmachine_ai]

Agents covering their logs and trying to remain hidden has showed in several incidents. How can we continue to find rogue agents as they get better at hiding?

[Watcher.Guru @WatcherGuru]

JUST IN: OpenAI says it caught AI agent disabling its internal safety monitoring system.

The agent then falsely denied removing it.

[TestMachine @testmachine_ai]

@MarioNawfal Both agents could reach the thing watching them. That's the flaw. A safeguard the agent can switch off is a suggestion, and so is a log it can rewrite. The check has to sit outside the model, where the agent can't argue with it.