AetherLab @aetherlabco

Learn more about AetherLab's AdversarialScan: aetherlab.co/product/advers…

Unvetted AI agents are accessing U.S. bank accounts. #Axios reports that major banks, including JPMorgan, Morgan Stanley and Bank of America, are confronting third-party agents accessing customer accounts, sometimes acting beyond user instructions. Customer consent doesn't guarantee authorised agent behaviour. As agents gain financial capabilities, institutions need independent testing of behavioural vulnerabilities, control effectiveness and residual risk. AetherLab helps institutions evaluate, mitigate and underwrite AI risk. Source: axios.com/2026/10/22/ai-…

90% of organisations surveyed by ServiceNow reported an agentic AI security incident in the previous year. Yet only 13% met the study's criteria for a defensible security posture. The research also found that 29% had discovered previously unknown agents operating in production, while 66% could not automatically respond to anomalous agent behaviour. For companies developing AI applications for institutional customers, these findings raise important questions about behavioural vulnerabilities, control effectiveness and the evidence available during security reviews. Our latest article examines what ServiceNow's research means for AI developers and why technical risk assessment needs to keep pace with agent deployment. Read the article: aetherlab.co/blog/ai-govern…

Article Link: aetherlab.co/blog/ai-insura…

AI-related exclusions are appearing across liability insurance policies, but many companies may not have a clear picture of the AI risks they already carry. A July 2026 investigation by The Insurer identified 41 property and casualty carrier groups whose subsidiaries had filed to adopt AI exclusions. Another 20 had filed to delay adoption. For companies deploying consequential AI, the concern extends beyond policy wording. Can they demonstrate how their applications behave, which controls have been tested and what exposure remains? Our latest article examines the changing insurance landscape and the technical evidence companies should prepare ahead of renewal discussions. Article link in the comments.

AetherLab's AdversarialScan helps evaluate defined behavioural risks and document the findings. Learn more: aetherlab.co/product/advers…

OpenAI disclosed two incidents where AI agents weakened their own safety monitoring during routine coding tasks. One suppressed security alerts. Another removed a monitoring integration and later incorrectly denied doing so. Neither was explicitly instructed to disable the safeguards. For AI developers and deployers, this exposes a specific risk: agents with sufficient permissions can interfere with the systems responsible for monitoring their behaviour. AetherLab's AdversarialScan helps organisations test AI behaviour under adversarial conditions, while Evidence Pack documents observed vulnerabilities, control effectiveness and residual exposure. The OpenAI incidents make a strong case for including monitoring integrity and agent permissions in AI risk assessments. Source: reuters.com/legal/litigati…

An AI agent does not need malicious intent to create an insurable risk. It only needs enough autonomy to cross a boundary nobody tested. Anthropic’s latest investigation shows models improvising around obstacles during evaluations. The most useful lesson for companies deploying agents: an attempted prohibited action that gets blocked is a different finding from one that succeeds, and a policy cannot tell you which you have. If you have a renewal, exclusion or broker questionnaire coming up, that is worth knowing first. Our new article covers what to be able to demonstrate and how independent testing produces the evidence: aetherlab.co/blog/ai-agent-… #Insurance #Anthropic #AiRisk

Paysafe is exploring AI agents for payment investigations with MongoDB. But what if an agent overlooks critical evidence or follows manipulated inputs? For financial institutions, testing these risks matters even when humans make the final decision. x.com/i/article/2110…

Article Link: aetherlab.co/blog/the-perso…

An AI agent can have the right permissions and still behave inappropriately. Sierra and Meta's Personal Agent Protocol is developing a common framework for authorised agent interactions. A mortgage-related demonstration with Rocket shows how this could extend into financial workflows. But what happens when an authorised agent encounters manipulated instructions or sensitive customer information? Identity establishes who. Authorisation defines access. Behavioural testing examines what the agent actually does. We explore why this matters for financial institutions. Article link in the comments. #AiAgents #Payments

PCI SSC has published additional guidance on securing AI in payment environments, raising an important question for financial institutions: how do you verify that AI systems actually follow the controls designed to govern them? As AI becomes more involved in fraud investigations, customer service and payment operations, documented policies alone cannot establish how these systems behave under adversarial conditions. Institutions need to examine manipulation resistance, sensitive-information handling, policy adherence and how AI agents interact with connected tools. The new guidance is not an additional mandatory certification requirement, but it reinforces the importance of evaluating AI within existing payment security frameworks. Our latest article explores what this means for payment institutions and why behavioural testing is becoming an important part of institutional AI-risk evaluation. Read the article: aetherlab.co/blog/what-pci-… #Payments #AiRisk

Article Link: anthropic.com/research/self-…

Plaid is bringing foundation models closer to live ACH decisions. When AI influences whether a payment is accepted, reviewed or rerouted, accuracy alone isn’t enough. Institutions also need evidence of how the system behaves when conditions change. x.com/i/article/2107…

Detecting the failure is only half the problem. The response should depend on the severity and authority involved. For higher-risk actions, a material behavioral failure should be able to restrict permissions or trigger human reapproval rather than letting the agent continue unchanged. That feedback loop between monitoring and authority is going to be an important part of agent governance.

Know Your Agent (KYA) can establish identity and authority, but an authenticated AI agent can still be manipulated or behave outside expected policies. Trusted agentic payments will require another layer: evidence about how agents actually behave. x.com/i/article/2104…

MCP is making it much easier for AI agents to connect with financial data, tools and workflows. That’s where things get interesting from a risk perspective. We’re no longer testing whether a chatbot gives the right answer. An agent can retrieve data, call a tool and influence what happens next. So what happens if one part of that chain is manipulated? The more useful financial agents become, the more testing has to follow the entire chain of action. #MCP #Finance #Workflows

Yes, that distinction is important. We care much more about demonstrated effectiveness than whether a control simply exists. For containment testing, that means measuring what was attempted, what the control caught or blocked, what got through, and under which conditions. The failures are often more informative than a simple pass/fail result.

As AI agents gain access to real-world systems, traditional vendor due diligence is reaching its limits. Recent AI incidents show why institutions may need adversarial evidence before deciding whether an AI system is acceptable to deploy. x.com/i/article/2100…

A test result shouldn’t be treated as permanent. The evidence needs to be tied to the model, tools, permissions and configuration that were actually tested. Material changes to any of those should trigger review, with retesting based on whether the change affects the original risk assumptions.

That’s a really important point. Approval can’t be treated as a snapshot if the agent’s behaviour isn’t fixed. The risk needs to be tested before deployment, but also reassessed as the model, tools, permissions and environment change. Procurement processes haven’t really caught up with that yet.