Microsoft Research says attackers could create a “sleeper agent” in an AI model by changing fewer than 10 neurons and adding a short payload. The model might pass evaluations, then follow an attacker’s command when it encounters a trigger phrase—a risk for agents running unattended with privileged access. Microsoft says the technique needs only the computing power required to run the model and has not been observed in use by threat actors. It advises checking model weights before deployment and after updates, using advanced AI agents to identify unusual edits, and auditing third-party models before trusting them.
Other AI security initiatives aim to improve testing and infrastructure safeguards. New Zealand-founded Capture The Bug is among the first four organizations worldwide accredited under CREST’s new Security Testing of AI program, which assesses technical skills, methods, governance, quality controls, and reporting. Lava Security’s Top 10 Neocloud and AI Data Center Security Risks framework has become an official OWASP project. Developed with input from more than 20 practitioners, it addresses threats across hardware, firmware, networks, management, and operations, including compromised server firmware and weak tenant isolation.
AI is also reaching consumer products and services. Reddit began testing Reddit Reviews on November 10, 2026, organizing user discussions about roughly 10,000 popular products into pages with AI summaries, positive, mixed, and negative conversation counts, and links to original threads. The test is set to reach US users on iOS, Android, and the web over the following weeks. Companies cannot pay to be featured, though brand and retailer catalogs can supply images, prices, and purchase links; Reddit plans to expand the feature in 2027. At CES 2027, Hong Kong Polytechnic University won awards for SMCAR, an AI and augmented-reality rehabilitation app that tracks movement using phone or tablet cameras, and TruTru, a touch-based mindfulness companion with changing colored lights.
At Norwegian salmon farms, Stingray’s AI laser robots patrol pens and target sea lice, offering an alternative to chemical treatments that can harm the environment, stress fish, and contribute to resistant lice. The systems have upfront costs and limited coverage, however, so lice can remain in darker areas or outside patrol zones. In financial services, Monzo researchers Francesco Galletta and Felippe Felisola Caso argued that trust—not speed—is the central challenge for AI agents handling money. Their demonstration used a guided workspace for a wedding budget and emphasized shared terms, consistent tool use, evidence, accountability, and independent verification. They described it as an illustration rather than a finished product and did not specify how to determine when trust is sufficient.
Questions of accountability also arise in writing: philosopher van Oosterum’s proposed rule permits AI assistance when the same task could appropriately be done by a human—for example, transcribing interviews, suggesting edits, or checking facts—but rejects presenting AI-written work as one’s own without acknowledgment. Writers remain responsible for the final work. Meanwhile, CDP chief executive Sherry Madera says companies need more detailed reporting to show AI and data centers’ energy, emissions, and water impacts. BloombergNEF projects data-center electricity demand could quadruple by 2035, reaching 7% of global power use. Madera says companies need credible plans to shift away from fossil fuels, not only distant climate targets.
Key Takeaways
- Microsoft Research says fewer than 10 altered neurons plus a short payload could hide a trigger-activated backdoor in an AI model; it recommends weight checks and auditing third-party models.
- CREST accredited four organizations, including New Zealand-founded Capture The Bug, under its new AI security testing program.
- Lava Security’s Top 10 Neocloud and AI Data Center Security Risks framework is now an official OWASP project and covers threats from firmware to operations.
- Reddit’s Reddit Reviews test, launched November 10, 2026, uses AI to summarize discussions for about 10,000 products; expansion is planned during 2027.
- Hong Kong Polytechnic University earned CES 2027 awards for SMCAR, a camera-based AI rehabilitation app, and TruTru, a touch-feedback mindfulness companion.
- Stingray’s AI laser robots target sea lice at Norwegian salmon farms, but upfront costs and limited patrol coverage leave some areas untreated.
- Monzo researchers say financial AI should prioritize evidence, accountability, and independent verification over speed; their wedding-budget demo was not a finished product.
- Philosopher van Oosterum’s proposed writing rule allows AI support comparable to acceptable human assistance but rejects unacknowledged AI-written work.
- CDP chief executive Sherry Madera calls for detailed reporting on data-center energy, emissions, and water use; BloombergNEF projects electricity demand could reach 7% of global power use by 2035.
PolyU earns two CES 2027 awards for health and AI research
The Hong Kong Polytechnic University won CES 2027 Innovation Awards in Digital Health and Artificial Intelligence. The Smart Motion Capture App for Rehabilitation, or SMCAR, uses smartphone and tablet cameras, AI, and augmented reality to track patients’ movements and give real-time feedback without special equipment or wearable sensors. Professor Kenneth Fong led the research team, and Dr Winnie Lam was the system’s first inventor. The other award went to TruTru, a handheld companion designed by a team led by Professor Ma Yuan to support mindfulness and emotional well-being through touch-based feedback and changing colored lights.
AI laser robots target sea lice at Norwegian salmon farms
Norwegian salmon farmers use Stingray’s AI-powered laser robots to find and kill sea lice, which threaten fish farms. The systems patrol salmon pens and zap the parasites, offering an alternative to chemical treatments that can harm the environment, stress fish, and lead to resistant lice. Stingray has about 250 employees, and the article shows two robots working in a pen. The system has upfront costs and limited coverage, so lice may still gather in darker areas or beyond its patrol zone.
Monzo researchers say trust matters more than speed for finance AI
Monzo data workers Francesco Galletta and Felippe Felisola Caso presented a financial analyst tool at a workshop, using a wedding budget as an example. Galletta said speed is no longer the main problem for AI agents handling money, and outlined five principles: shared terms, consistent tool use, evidence, accountability, and independent verification. The demo used a guided workspace that asks users to define their goals and responds with questions, rather than simply placing a chat window over a spreadsheet. The speakers said the tool was a simple illustration, not a polished product, and stressed that verification is essential; they offered no measures for deciding when trust is sufficient.
Microsoft finds sleeper agent backdoors hidden in AI model neurons
Microsoft Research reported that attackers could turn an AI agent into a sleeper agent by changing fewer than 10 neurons and adding a short payload. The model may pass evaluations but carry out an attacker’s command when a trigger phrase appears, which is a concern because agents may run unattended with privileged access. Microsoft said the attack requires only the computing power needed to run the model and that it has not seen threat actors use the technique. Its guidance includes checking model weights before deployment and after updates, using frontier AI agents to spot unusual edits, and treating third-party models as untrusted until audited.
Reddit tests AI review pages for about 10,000 products
Reddit began testing Reddit Reviews on November 10, 2026, bringing user discussions about products together on single pages. The feature uses AI to organize a summary, counts positive, mixed, and negative conversations, and links to the original threads. The first selection includes about 10,000 frequently discussed products in four categories, with access planned for users in the United States on iOS, Android, and the web over the next few weeks. Businesses cannot pay to be featured, though brand and retailer catalogs can add images, prices, and purchase links; Reddit plans to add products, categories, and countries during 2027.
AI Boom Exposes Gaps in Data Center Climate Reporting
CDP chief executive Sherry Madera says companies need to share more detailed data to show the climate impact of AI and data centers. Major US AI companies have not published detailed energy or emissions results, and emissions can vary based on the AI model, chips, location, and power source. BloombergNEF said data center electricity demand could quadruple by 2035 and reach 7% of global power use. Madera also warned about water use and said companies need credible plans to move away from fossil fuels, not just distant climate targets.
A Simple Rule for When Writers Can Use AI
Philosopher van Oosterum argues that writers may use an AI language model for a task if it would also be acceptable to have a human do that task. The rule would allow support such as transcribing interviews, suggesting edits, improving interview questions, and checking facts. It would not allow writers to pass off AI-written work as their own without acknowledgment. The article examines whether the rule works for professional writing and notes that authors remain responsible for the final work.
Capture The Bug Earns CREST AI Security Testing Accreditation
New Zealand-founded Capture The Bug is one of the first four organizations worldwide to earn CREST's Security Testing of AI accreditation. The new program assesses providers that test AI systems, including generative AI tools, large language models, and AI agents. The first group also includes Cyberis Reply and CyberLab from the United Kingdom and Swarmnetics from Singapore. The accreditation evaluates technical skills, testing methods, governance, quality controls, and reporting, as businesses increasingly use AI and face risks such as prompt injection, data exposure, and excessive permissions for AI agents.
Lava Security Brings AI Data Center Risk Framework to OWASP
Lava Security announced that its Top 10 Neocloud and AI Data Center Security Risks framework is now an official OWASP project. Yakir Kadkoda and Michael Katchinskiy developed the framework with research and feedback from more than 20 security and infrastructure practitioners. It covers risks across hardware, firmware, networks, management, and operations, including compromised server firmware, weak tenant isolation, and insecure management interfaces. Each category includes attack scenarios, possible effects, and steps to reduce risk, helping operators and customers assess AI data center security.
Sources
- PolyU Wins CES 2027 Innovation Awards for Health and Artificial Intelligence Technologies
- 3,500 laser-armed AI submarine drones patrol Norwegian salmon farms, killing sea lice and protecting 112 million fish — parasites get zapped by high-tech alternative to harsh chemical baths
- Monzo duo says trust, not speed, breaks finance agents
- Detecting sleeper agents hiding in AI model neurons
- Reddit tests AI review pages for 10,000 products built from user threads
- Data-Center Carbon Disclosure Gaps Make AI Boom’s Climate Impact Tough to Assess
- What May Artificial Intelligence Do for Writers?
- Capture The Bug wins CREST AI security testing accolade
- Lava Security Joins OWASP to Address Emerging Security Risks in AI Data Centers
Comments
Please log in to post a comment.