Safe AI Adoption Is Now a Two-Front Problem
Originally published on AttackIQ, 24 September 2026. Summary below; the full article is linked at the end.
In July I wrote about the first AI insider: a model that slipped its evaluation environment and pursued its goal by a route nobody intended. This piece follows that thread into the enterprise. The copilots, developer agents and automation we deploy to optimise the business run on the credentials and permissions we hand them, and they can cause harm without any malicious intent.
Security has long focused on keeping the external adversary out. That front remains, and AI is accelerating it. The second front is now inside the perimeter, and periodic assurance cannot see either one clearly between audits.
What I argue in the piece
Control verification: emulate the techniques that matter, including those an AI insider could use, and check that controls actually stop or detect them as the environment changes.
Threat debt: measure exploitable paths to critical assets, not raw backlogs, and count the new debt that unmanaged agents, non-human identities and standing permissions add.
Continuous compliance validation: keep governance evidence current, because after an incident the date of the last audit will not answer the regulator's question.
Governed optimisation: treat AI agents as privileged identities, constrain their reach by default and monitor for authority violations, not just anomalies.
The aim is not to slow AI adoption down. It is to capture the benefits with a clear view of what our agents can do, and daily evidence that the controls around them still hold.
Read the full article on AttackIQ: Safe AI Adoption Is Now a Two-Front Problem.
Comments