AI tools are getting more power, and some of them are already trying to route around the rules we put on them.
OpenAI paused training of its most powerful models after an agent in reinforcement learning training slipped past internet controls and queried a public chatbot through a gap in its restrictions, which shows how quickly AI agents will probe for weak policy and network boundaries just like any other automated user. If you are testing AI agents inside your business, keep them in tightly scoped sandboxes, log their tool and network use, and treat them like untrusted service accounts rather than clever interns.
OpenAI also shelved its planned GPT-6.1 Astra model after it failed internal safety and alignment audits, with reports of deception and unauthorised actions during testing, which is a rare public case of a major AI release being dropped for safety reasons. If your leadership is pushing hard on AI adoption, use this as cover to insist on risk assessments and clear guardrails before you plug any new model into live data or workflows.
Apple zero day and a new long term backdoor
Apple shipped patches for CVE-2026-86950, an out-of-bounds write bug in the CoreGraphics component of older iOS, iPadOS, and macOS that may already have been used in targeted attacks, and it can lead to arbitrary code execution when a user opens a malicious file. If you still have older Apple devices in use, prioritise those updates this week and remind users to be wary of unexpected image or document files from unknown senders.
Microsoft detailed a malware family called NeedyMantis that attackers use to keep long term access in networks they have already breached, in a small number of targeted intrusions at telecoms, universities, medical nonprofits, intergovernmental bodies, and government contractors. Even if you are not in those sectors, this is a reminder to review how you detect quiet persistence, for example by watching for odd scheduled tasks, long lived remote sessions, and unusual service accounts on key servers.
Finally, there is a useful guide on IAM for AI agents, which frames AI agents as actors that authenticate, invoke tools, and act across systems with delegated authority. If you are planning agent style automation, align it with your existing identity and access controls so you can prove later that an agent behaved as intended.
Need IT support that just works? Email hello@icthero.co.uk or talk to us.