Research & Blog

Notes from the adversarial side.

Writing on AI risk, adversarial testing, and governance, from the people red-teaming production AI systems every week.

AllFuture PredictionsDeep TechGovernance & PolicyRed-Team Field NotesPaymentsInsurance
AI agent risk assessment highlighting Anthropic's findings on unintended agent behaviour, security boundaries and evidence for AI liability insurance.
INSURANCEOct 16, 2026 · 5 min read

AI Agent Risk: What Anthropic's Findings Mean for Insurance Evidence

Anthropic's investigation documented AI agents exploiting a software vulnerability, submitting unintended online forms and finding alternative routes around restrictions. For companies facing AI-related insurance exclusions, renewals or underwriting questions, these findings highlight why policies alone are insufficient. Independent behavioural testing can provide evidence of attempted actions, control effectiveness and unresolved risks.

Anders MogensenRead →
Personal AI agent interacting with a financial institution's digital systems, illustrating identity, authorisation, behavioural testing and operational risk.
PAYMENTSOct 14, 2026 · 5 min read

The Personal Agent Protocol Is Taking Shape. What Does It Mean for Financial AI Risk?

Sierra and Meta's Personal Agent Protocol is introducing a common framework for interactions between personal AI agents and businesses. With financial institutions participating in its development and an early mortgage-related demonstration, the initiative raises a critical question: how can institutions evaluate the behaviour of agents operating within legitimate permissions?

AetherLabRead →
Illustration of an AI-powered payment infrastructure with connected financial systems, representing PCI SSC guidance on AI security, adversarial testing and institutional risk assessment.
PAYMENTSOct 13, 2026 · 9 min read

What PCI's New AI Guidance Means for Payment Institutions

PCI SSC's new guidance highlights the growing security challenges of AI in payment environments. For banks, payment processors and financial institutions, the next step is understanding how AI systems behave when instructions are manipulated, sensitive information is exposed or operational boundaries are challenged.

Alex Georges, PhDRead →
TECHNICALJul 8, 2026 · 7 min read

Read the NIST Guardrail Proof as a Liability Document

A formally correct extension of Gödel to AI guardrails tells attackers and defenders nothing they didn't already know. Its real payload is what it does to approval paper trails, and the math worth studying is in how guardrail failure actually scales.

Alex Georges, PhDRead →

Reading because you have an AI system to approve?

Skip the archive. Tell us what you are deploying, approving, or underwriting, and we will scope an assessment against your break-goals.

Request an AI risk assessment

Get new writing by email.

Occasional notes on AI risk, adversarial testing, and governance. No marketing drip.