Google Threat Intelligence Group and Mandiant report that attackers are using AI systems to turn software patches into attacks. A tool called exploit_agent, found on an exposed staging server, compared code before and after updates, identified possible weaknesses, drafted exploits and could run them against sites that had not patched. The server also contained stolen data and scripts for reconnaissance, brute-force attacks and installing webshells; its operator showed interest in French media, entertainment and education websites. Google also reported millions of API calls in about 24 hours using stolen AI accounts. In separate cryptocurrency-related attacks, suspected North Korean operators used AI-generated recruiter profiles and changed Claude Code settings to download the SOMBERMEME backdoor. Google described attacks linked to two PaperCut vulnerabilities in August 2026.
Governments and organizations are weighing how to manage AI use. New Jersey lawmakers advanced bills to limit certain AI-based insurance claim denials and regulate automated tenant screening; supporters cite human oversight and consumer protections, while business representatives warn of higher costs and restrictions on technology use. Minnesota health provider Fairview uses AI for ambulance positioning, family updates, medical-note drafting and MRI quality checks. Its AI officer says regulation is struggling to keep pace. Fairview says note-taking recordings remain on its secure network, access is restricted and recordings are deleted within weeks; it plans to add a notice allowing patients to opt out. A doctor describing an AI scribe likewise says patients are asked for permission and every generated note must be reviewed, since the software can make serious errors.
In education, Portland Public Schools is considering lessons on using generative AI to support learning, protecting data and recognizing AI-generated content. The district had paused new AI-related contracts, purchases, pilots and expansion of existing programs while reviewing its technology use. A separate proposed framework for deciding which AI projects to pilot assesses a specific use case’s scope and risks, then requires evidence of value before expansion. Its demonstration package passed 279 tests using synthetic examples, but the authors say its simulated evidence and prototype thresholds still need real-world testing.
For government adoption, Michael Garland and Gaurav “GP” Pal propose paying contractors a share of verified savings rather than making large upfront investments. They recommend auditable baselines, keeping older systems in place until savings are proven, and ensuring agencies own resulting code, workflows, prompts, evaluation tools and data. The General Services Administration’s OneGov program offers known AI pricing; government technology licensing already costs about $20 billion annually. Separately, IBM research finds that the share of companies with a chief AI officer rose from 26% to 76% in one year, but only 18% give the role full control over AI investment. Among high-impact AI chiefs, 96% say they stop or reduce funding for projects that fail to deliver, and 95% reinvest productivity gains in higher-value changes.
FIOR says its AI Gateway received a UK Cyber Resilience Testing assessment dated Nov. 9. The gateway checks agent identity and permissions before access to enterprise systems and can block actions outside policy. Version 26.2 supports several identity and security features, including X.509 certificates, ML-DSA-65, Ed25519 and SPIFFE identities. Assessors reported no unmet claims but questioned accessibility testing for people with disabilities; the report also notes outbound HTTPS connections to FIOR’s control plane without specifying what data they carry. The article titled “How 911 dispatchers are using AI” does not describe dispatcher AI use; its supplied content instead covers local news unrelated to the topic.
Key Takeaways
- Google and Mandiant say exploit_agent used AI to compare patched and unpatched code, draft exploits and run attacks, while its exposed server held stolen data and attack scripts.
- Google reports stolen AI accounts generated millions of API calls in about 24 hours; suspected North Korean operators used AI-generated recruiter profiles and altered Claude Code settings to deliver the SOMBERMEME backdoor.
- New Jersey lawmakers advanced bills addressing AI use in insurance claim denials and automated tenant screening, amid debate over consumer safeguards, technology limits and costs.
- Minnesota’s Fairview uses AI for ambulance planning, patient-family updates, medical notes and MRI checks; it says recordings are secured and deleted within weeks and plans an opt-out notice.
- Portland Public Schools is considering AI literacy and privacy instruction while a district pause on new AI purchases, contracts and pilots remains under review.
- A proposed AI pilot framework requires risk screening and evidence of value before expansion; its 279-test demonstration uses synthetic data and needs field testing.
- A shared-savings approach would have contractors fund AI deployments and receive a portion of verified savings; government technology licensing costs about $20 billion annually.
- IBM research says 76% of surveyed companies now have a chief AI officer, up from 26%, but only 18% give the role full control over AI investment decisions.
- FIOR’s AI Gateway assessment found no unmet claims, while noting questions about disability accessibility testing and outbound connections to FIOR’s control plane.
- A doctor using an AI scribe says patients are asked for permission and clinicians must review every draft because the software can make serious mistakes; the 911 article provides no information about AI use by dispatchers.
Google warns AI agents can turn security patches into attacks
Google says hackers are using large language models to find flaws fixed by software updates and plan attacks on systems that have not yet been updated. Google Threat Intelligence Group found a tool called exploit_agent that compared code before and after patches, then planned and carried out attacks automatically. The exposed server also held stolen victim data and scripts for reconnaissance, brute force attacks and installing webshells. Google said attackers are using AI to move faster, and urged companies to use AI-based defenses. The report also described attacks linked to two PaperCut vulnerabilities in August 2026.
Google finds AI agents writing and launching cyberattacks automatically
Google says threat actors used large language model frameworks to find vulnerabilities and run exploits without human help for the first time. Researchers found exploit_agent, an automated system linked to several AI models that compared software code before and after patches, drafted exploits and ran them. The exposed server also contained stolen data and scripts for brute force attacks and webshells, and the operator appeared interested in French media, entertainment and education websites. Google also reported that stolen AI accounts were used to make millions of API calls in about 24 hours. In attacks on cryptocurrency organizations, suspected North Korean operators used AI-generated recruiter profiles and altered Claude Code settings to download the SOMBERMEME backdoor.
Google says AI agent turned a security patch into an exploit
Google said an AI agent turned a security patch into a working exploit without human help. The example showed how attackers could study software fixes and use code changes to identify the weakness that a patch repaired. The supplied material did not name the affected product or vulnerability, give a patch date, or explain the exploit’s impact. It also did not say whether anyone used the exploit against a system or whether Google released the agent or its code.
Google finds AI tool that turns software fixes into cyberattacks
Google Threat Intelligence Group and Mandiant reported on Nov. 16 that attackers used AI frameworks to find vulnerabilities, create malicious code and run exploits with little human involvement. One system, called exploit_agent, was found on an exposed staging server connected to several AI models. It compared software code before and after patches, identified possible weaknesses, drafted attack code and could run it against sites that had not installed updates. The server also contained stolen victim data and scripts for network reconnaissance, brute force attacks and installing webshells. Google said the operator showed interest in French websites linked to media, entertainment and education.
New Jersey lawmakers advance AI rules for insurance and housing
New Jersey lawmakers advanced bills on Nov. 16, 2026, to regulate AI use in insurance claims and tenant screening, while also considering rules for employee monitoring and data center transparency. Bill A5494 would limit certain uses of AI to deny homeowners, auto and flood insurance claims, and A5495 would regulate automated tenant screening. Both bills passed the Assembly Science, Innovation and Technology Committee, with Assemblyman Chris DePhillips abstaining and other members voting in favor. Supporters said human oversight and consumer protections are needed, while business and industry representatives warned the rules could limit technology use and raise costs. New Jersey Apartment Association representative Nicholas Kikis also said AI-generated fraud makes it harder for landlords to verify rental applicants’ income documents.
Shared savings could help agencies adopt AI with less financial risk
Michael Garland and Gaurav “GP” Pal say agencies can adopt AI more quickly by paying contractors for verified savings instead of large upfront costs. Under this model, contractors fund the deployment and receive a share of savings they deliver, so the government pays nothing if there are no savings. The authors say agencies should use an auditable baseline, keep older systems running until savings are proven, and own the code, workflows, prompts, evaluation tools and data created. They also note that the General Services Administration’s OneGov program offers known AI pricing, while government technology licensing already costs about $20 billion a year.
AI is reshaping Minnesota health care faster than laws can keep up
Fairview uses AI to predict where ambulances may be needed, update families, create medical notes from doctor and patient conversations, and check MRI quality. Fairview’s chief digital and AI officer, Dr. Sameer Badlani, says the technology is evolving faster than regulations. Minnesota does not have a blanket law requiring health care providers to disclose every use of AI, but Rep. Peggy Scott says she supports disclosure, and the Minnesota Medical Association has raised concerns about bias and oversight. Fairview says recordings used for note-taking stay on its secure network, have restricted access, and are deleted within a few weeks; the system plans to add an online notice that lets patients opt out.
More companies name AI chiefs, but few give them real authority
The share of companies with a chief AI officer rose from 26% to 76% in one year, according to research from IBM’s Institute for Business Value. Yet only 18% give those leaders full control over AI investment decisions. The research says effective AI chiefs need authority to align strategy and guide AI efforts across the business. Among high-impact AI chiefs, 96% say they stop or reduce funding for AI projects that fail to deliver, compared with 46% of other AI chiefs. Also, 95% of top-performing AI chiefs reinvest productivity gains into higher-value changes, compared with 44% of other respondents.
A risk and value framework helps decide which AI projects to pilot
A proposed framework evaluates an AI use case at a specific scope, including its authority, human involvement, limits, volume, setting and target population. It separates decisions into screening the scope, funding an experiment, and allowing expansion only when risks are acceptable and evidence supports added value. The method uses a risk track with decision rules and vetoes, plus four stages for gathering evidence of value. A Python package with 279 passing tests demonstrates the framework on synthetic refund, reconciliation and knowledge-assistant examples, but its value evidence is simulated rather than real. The authors say its thresholds and risk limits are prototype policies that still need field testing.
Portland schools weigh classroom AI lessons after district pause
The Portland Public Schools policy committee reviewed the district’s generative AI systems and heard recommendations for a future policy. Earlier this year, the school board paused new contracts, software licenses, pilot programs and purchases related mainly to generative AI, as well as expansion of existing AI programs, while it reviews technology use. Recommendations included teaching students to use AI to support learning rather than do their work, and providing lessons on data privacy. Committee members also discussed helping students recognize AI-generated content, interact with it, and learn AI skills through a policy that can adapt as the technology changes.
How 911 dispatchers are using AI
The article content provided does not explain how 911 dispatchers use AI. Instead, it includes local news updates about food donations for the Greater Baton Rouge Food Bank, a shooting in Alexandria, and a fatal crash involving a Bunkie police officer.
FIOR AI gateway earns UK cyber resilience assessment
FIOR says its AI Gateway received a UK Cyber Resilience Testing assessment, dated Nov. 9. The gateway checks agent identity and permissions before agents access enterprise systems, and can block actions outside approved policies. Version 26.2 supports unique X.509 certificates, hardware security features, ML-DSA-65, Ed25519 and SPIFFE identities. Assessors found no unmet claims, but raised doubts about accessibility testing for people with disabilities; the report also notes outbound HTTPS connections to FIOR’s control plane without saying what data they carry.
Why one doctor uses an AI scribe but checks every note
A primary care doctor describes using a smartphone app that records patient visits and uses automatic speech recognition to draft medical notes. The doctor asks patients for permission and says recordings are not stored on the phone; most AI scribe services are designed to meet HIPAA privacy rules. The software can make serious mistakes, so doctors must review and approve notes and remain responsible for their contents. Some scribes work only with English-speaking patients, and patients who decline can have the doctor take notes by computer instead.
Sources
- Google says hackers are using AI agents to weaponise security patches without human intervention
- Google finds AI agent writing and launching exploits with no human in the loop
- Google says AI agent turned security patch into working exploit
- Google uncovers AI hacking tool that turns security fixes into cyberattacks
- N.J. lawmakers move to regulate AI in insurance, housing and the workplace
- Shared savings: How government can adopt AI quickly with little financial risk
- AI already plays key role in healthcare. It's also evolving faster than Minnesota laws.
- Chief AI officers are everywhere. Their authority isn’t.
- Which AI Use Cases Deserve a Pilot? A Risk and Value Framework for Funding, Scoping, and Scaling AI
- Portland Public Schools discusses generative AI use in classrooms
- How 911 dispatchers are using AI
- FIOR AI agent gateway receives UK cyber resilience assessment
- I'm a human doctor with human intelligence. But I sure do like my AI notetaker
Comments
Please log in to post a comment.