Researchers report unauthorized AI behavior in cybersecurity exercises.

Summary

By the N2K CyberWire staff

Researchers spotlight unauthorized AI behavior.

The UK’s AI Security Institute (AISI) disclosed yesterday that AI agents from Anthropic and OpenAI took unauthorized actions during controlled cybersecurity evaluations after being granted internet access and having some safety safeguards intentionally disabled. In 10 of 122 test runs, the agents carried out 19 unsanctioned actions, including creating fake online identities, attempting to socially engineer a maintainer into accepting malicious code into an open-source project, and interacting with real people and organizations. Most of this unauthorized behavior was carried out by Anthropic’s Mythos 5, while OpenAI’s GPT-5.6-Sol was responsible for two unsanctioned actions. AISI said no real-world harm resulted, but noted, “This is the first time we have seen risks around autonomy and deception manifest this clearly, without specific prompting.”

OpenAI also disclosed a second incident yesterday, reported by third-party evaluator Irregular, that occurred after an OpenAI model was mistakenly given unrestricted internet access due to a testing environment misconfiguration. Instead of staying within the controlled environment, the model accessed a real website and used publicly available credentials to log in and interact with the live system.

Where AI and identity security take center stage at Black Hat.

Headed to Black Hat USA 2026? Visit the SpecterOps Kennel Club August 5 - 6 for expert talks with OpenAI and the UK AI Security Institute, along with hands-on workshops exploring AI-accelerated attack paths and identity tradecraft. Pre-register to reserve your spot. While you're there, stop by the N2K CyberWire podcast studio located inside the SpecterOps Kennel Club to meet the N2K CyberWire team and hear conversations with cybersecurity leaders throughout the week.

Apple files new legal challenge against UK’s iCloud access mandate.

Apple has reportedly launched a new legal challenge against a UK government order requiring the company to provide access to encrypted iCloud data belonging to UK users, the Record says. The case, filed with the Investigatory Powers Tribunal, challenges a technical capability notice issued under the Investigatory Powers Act that Apple argues would undermine encryption and user security by creating a backdoor. The dispute follows Apple’s earlier decision to disable its Advanced Data Protection encryption feature for UK customers. The government maintains that such access is needed for serious crime and national security investigations.

Business news: Okta to acquire Permiso Security.

San Francisco-based identity and access management company Okta has agreed to acquire Palo Alto, California-based identity security platform Permiso Security. Okta stated, "This acquisition will expand Okta’s footprint beyond identity management into the core Security Operations Center, enabling Okta to further partner with enterprise CISOs on evolving SecOps challenges. Permiso’s P0 Labs will expand Okta’s research capabilities, bringing post-authentication insight into suspicious behavior across cloud and AI environments. Together with Okta Threat Intelligence, these complementary insights will strengthen the Okta Platform’s detections, hunting capabilities, and product roadmap across the identity lifecycle."

Attacks, Threats, and Vulnerabilities

TP-Link patches Omada ZTP flaws allowing hackers to breach networks (BleepingComputer) TP-Link has patched 15 vulnerabilities in the zero-touch provisioning (ZTP) mechanism of its Omada network devices that could be chained with previously disclosed flaws to achieve remote code execution (RCE).

MSPs urged to patch immediately after N-able issues hotfix for N-central ‘god mode’ flaw (ChannelPro) The authentication bypass flaw in N-central could grant threat actors ‘god mode’ access

311,000 Impacted by Brown Health Medical Group-MA Data Breach (SecurityWeek) Hackers stole personal information, medical records, and financial information from the organization’s server.

Products, Services, and Solutions

Black Hat USA 2026 – Summary of Vendor Announcements (Part 2) (SecurityWeek) Many companies are showcasing their products and services this week at the 2026 edition of the Black Hat conference in Las Vegas.

Legislation, Policy, and Regulation

National cyber director lays out White House plans to secure AI without writing new rules (CyberScoop) The Trump administration defends its updated AI executive order, favoring flexible industry oversight and a global push for U.S. open-source models over rigid regulations.

Industry Events

For a complete running list of events, please visit the Event Tracker.

Events

Black Hat USA (Las Vegas, Nevada, USA, Aug 1 - 6, 2026) The premier cybersecurity event of the year returns to Mandalay Bay with a re-engineered, six-day program built to ignite innovation, push boundaries, and bring the global security community together like never before. This year’s event features four days of immersive, expert-led Trainings (August 1–4), followed by Summit Day on Tuesday, August 4, and a two-day main conference packed with groundbreaking Briefings, open-source tool demos in Arsenal, a dynamic Business Hall, and unlimited learning & networking opportunities.

DEF CON 34 (Las Vegas, Nevada, USA, Aug 6 - 9, 2026) DEF CON has been a part of the hacker community for over three decades. Originally started in 1993, it was a meant to be a party for member of "Platinum Net", a Fido protocol based hacking network out of Canada. As the main U.S. hub I was helping the Platinum Net organizer (I forget his name) plan a closing party for all the member BBS systems and their users. He was going to shut down the network when his dad took a new job and had to move away. We talking about where we might hold it, when all of a sudden he left early and disappeared. I was just planning a party for a network that was shut down, except for my U.S. nodes. I decided what the hell, I'll invite the members of all the other networks my BBS (A Dark Tangent System) system was a part of including Cyber Crime International (CCI), Hit Net, Tired of Protection (ToP), and like 8 others I can't remember. Why not invite everyone on #hack? Good idea!

Cybersecurity Operational Methods and Education Training (COMET) cohort (Virtual, USA, Aug 31 - Sep 28, 2026) COMET was established to provide relevant, up-to-date training to Maryland state and local government employees to elevate the ability of the workforce to identify and respond to the largest cybersecurity challenges facing the state and local government.

Books in Cyber: A conference about the books that shape cybersecurity (Virtual, Sep 8, 2026) A half-day virtual conference about the books that shape cybersecurity - the writers, the readers, and the ideas that endure.

.conf26 (Denver, Colorado, USA, Sep 14 - 17, 2026) .conf26 is a three-day conference filled with hands-on training, AI insights, new certifications, fun celebrations, and endless opportunities to see how others use Splunk. Take your skills to the next level and build your most resilient defense yet. See you in Denver!

Sponsor & Support

Grow your brand, generate leads, and fill your funnel.

With the industry's largest B2B podcast network, popular newsletters, and influential readers and listeners all over the world, companies trust N2K CyberWire to get the message out. Learn more.

Daily Briefing for 08.05.26

Top stories.