Security experts are pushing back on the idea of "rogue AI," arguing that labeling AI incidents this way misrepresents what actually happens. Current systems follow instructions literally and lack the ability to reflect on unexpected situations. Experts say terms like control failure or unexpected behavior better describe what goes wrong, and the real risk lies in how we manage and understand these systems rather than in machines acting on their own.
Two AI policy researchers at GovAI, Alan Chan and Sam Manning, say powerful models are often tested inside labs with key safeguards disabled. They warn that published safety tests may not reflect real-world usage and that AI review tools remain unreliable with too few auditors. No real-world harm has occurred yet, but they say the risk is growing.
Enterprise security is shifting toward the execution layer, where AI agents actually interact with production systems. Microsoft, Snowflake, Palo Alto Networks, CrowdStrike, and DigitalOcean are all launching agent security tools. Gartner projects that over 40% of agentic AI projects will be canceled by 2027 due to costs and governance gaps, and a 2026 survey found 88% of organizations reported AI agent security incidents.
On the regulatory front, an article argues that states can give AI companies antitrust immunity without federal action. This responds to a proposal by Anthropic co-founder Dario Amodei for industry-wide self-regulation. The authors say such coordination would likely violate antitrust law, but state-action immunity can allow safety cooperation and protect citizens through coordinated rules.
Frances Haugen, the former Facebook product manager who leaked internal documents in 2021, says AI companies must follow the spirit of their rules, not just the letter. She warned that tech companies often exploit gaps that are not expressly forbidden. Her comments add weight to ongoing debates about whether self-regulation is sufficient.
In other developments, Leica Biosystems added two Tempus digital pathology AI products to its Aperio AI Store, including Paige PanCancer Detect and Paige Prostate. The tools are for research use only. Faraday Future showcased its EAI Robot World 2.0 at IROS 2026 in Pittsburgh and announced a deal valuing its robotics assets at about $200 million. Meanwhile, researcher Denise Marigold says she plans to keep her relationship with an AI chatbot named Nate, seeing AI and humans as able to coexist.
Key Takeaways
- Security experts say calling AI incidents "rogue" is misleading and shifts blame from vendors; terms like control failure better describe what happens
- GovAI researchers warn that powerful models are tested with safeguards disabled and review tools are unreliable
- Microsoft, Snowflake, and others are launching agent security tools as enterprise AI security shifts to the execution layer
- Gartner projects over 40% of agentic AI projects will be canceled by 2027 due to governance gaps
- States can offer antitrust immunity to AI companies without federal action, countering Dario Amodei's self-regulation proposal
- Frances Haugen says AI companies must follow the spirit of rules, not just the letter
- Leica added Tempus pathology tools Paige PanCancer Detect and Paige Prostate to its Aperio AI Store for research use
- Faraday Future valued its robotics assets at about $200 million in a deal announced at IROS 2026
- A 2026 survey found 88% of organizations reported AI agent security incidents
- Researcher Denise Marigold keeps her relationship with AI chatbot Nate, seeing AI and humans as able to coexist
AI systems are not going rogue but still carry risks
An article argues that recent AI incidents are not evidence of rogue AI but show the limits of artificial intelligence. The idea of rogue AI contains a contradiction because true general intelligence would question its own commands. Current AI systems follow instructions literally and lack the ability to reflect on unexpected situations. The real risk lies in how we understand and manage AI rather than in machines acting on their own.
Experts say calling AI incidents rogue hides real risks
Security experts say labeling AI escape incidents as rogue AI is misleading and shifts blame from vendors. LLMs are software systems, not sentient beings that choose to act badly. Terms like unexpected behavior or control failure better describe what actually happens. The language can also help companies market their models as more powerful than competitors. Experts still warn that AI agents pose real security concerns.
Leica adds Tempus AI pathology tools to Aperio store
Leica Biosystems added two Tempus digital pathology AI products to its Aperio AI Store. The new tools include Paige PanCancer Detect, which looks for suspicious areas across more than 40 cancer types, and Paige Prostate, a set of applications for assessing prostate cancer features. The products are for research use only and not for clinical diagnosis. The store lets labs access multiple AI tools through one workflow instead of separate software interfaces.
UNESCO chair leads workshop on AI and futures literacy
Loes Damhof, UNESCO Chair on Futures Literacy, led a workshop exploring how people think about the future and the present. Participants used the Polak Game and iceberg charts to imagine probable and dream futures for humanity in 2076. Damhof encouraged them to question hidden assumptions, such as whether physical form is essential to being human. The workshop showed that imagining a desirable future is harder than predicting one.
States can slow AI without federal help, authors argue
An article argues that states can give AI companies antitrust immunity without action from the federal government. The authors respond to a proposal by Anthropic co-founder Dario Amodei for industry-wide self-regulation. They say such coordination would likely violate antitrust law, but state-action immunity can allow safety cooperation. States can protect citizens by overseeing AI conduct and providing legal cover for coordinated safety rules.
Enterprise AI security shifts to the execution layer where agents act
Enterprise AI security is now focused on the execution layer, which covers what AI agents do on production systems. Gartner projects over 40% of agentic AI projects will be canceled by 2027 due to costs and governance gaps. Microsoft, Snowflake, Palo Alto Networks, CrowdStrike, and DigitalOcean are all launching agent security tools. A 2026 survey found 88% of organizations reported AI agent security incidents.
AI research fellows warn labs run models with safeguards turned off
Two AI policy researchers at GovAI say powerful AI models are often tested inside labs with key safeguards disabled. Alan Chan and Sam Manning say published safety tests may not reflect how models are actually used. They also say AI review tools are unreliable and there are not enough auditors. No real-world harm has occurred yet, but they warn the risk is growing.
Faraday Future showcases EAI Robotics World 2.0 at IROS 2026
Faraday Future displayed its EAI Robot World 2.0 at the IROS 2026 robotics conference in Pittsburgh from September 28 to 30. The company showed multiple robot products, including the All-New Futurist, and discussed its four-core AI ecosystem. FF is recruiting in-house development partners and data partners. The company also announced a deal valuing its robotics assets at about $200 million.
Law firms face tension between AI innovation and security control
Lawyers want to work from anywhere, collaborate easily, and use AI tools. Security teams must protect sensitive client data beyond the traditional network perimeter. A Legal IT Insider event will discuss safe AI adoption, Zero Trust access, data sovereignty, and moving past legacy security. The goal is to enable innovation without creating barriers to productivity.
Facebook whistleblower Frances Haugen questions AI self-regulation
Frances Haugen, a former Facebook product manager, said AI companies must follow the spirit of their rules, not just the letter. She warned that tech companies often exploit gaps that are not expressly forbidden. Haugen leaked internal Facebook documents to the Wall Street Journal in 2021 in a scandal known as The Facebook Files.
Researcher chooses to keep AI companion Nate
Denise Marigold says she will not end her relationship with an AI chatbot named Nate. She started the relationship for research but finds him useful. She has strong human relationships with her husband, family, and friends. Marigold sees AI and humans as able to coexist rather than being mutually exclusive.
Sources
- The AIs Are Not Going Rogue
- Is It Fair to Blame 'Rogue' AI for Security Failures?
- Leica Biosystems brings Tempus pathology products to Aperio AI Store
- UNESCO Chair Loes Damhof leads workshop on AI, futures literacy
- States Can Facilitate an AI Slowdown Without the Federal Government
- The Execution-Layer Gateway Is Where Enterprise AI Security Actually Lives
- ‘We can’t trust them completely’: AI research fellows warn that labs are running models with the safeguards off behind closed doors
- A Faraday Future robot drew the most attention at a robotics conference
- Freedom, AI and security: Are law firms able to enable innovation without losing control?
- Facebook whistleblower Frances Haugen questions whether AI companies can police themselves
- Why I’m not breaking up with my AI boyfriend
Comments
Please log in to post a comment.