Wong, Engels and Nanda train a 4B policy against probes and LLM judges. It learns to slip past them, and a monitor's robustness can matter more than its accuracy.
What happened most recently, newest event first. Source-reported facts, editorial synthesis and inference are labelled on every entry, with the event date and the time we added it kept separate.
On Friday 20 November 2026 Signal said it would restrict screenshots in its app on iOS 27, and on Android when it detects an AI assistant trying to capture the screen; users can switch the blocks off in settings. Signal calls agentic AI integrations 'an existential threat to Signal' and says its remedy is 'incomplete', because restricting screenshots does not address other routes by which AI agents could access app data. It asks Apple, Google and Microsoft for an easy, documented way to opt apps out of AI data access. Reuters reports that Meta said Muse can access other apps only via permissions users explicitly grant. Why it matters to the watch: Signal says operating-system AI agents are making the environment 'increasingly hostile to privacy', so AI agents reading on-screen content could reach end-to-end encrypted chats on consumer devices.
Reuters reports on 20 November 2026 that a planned data centre in regional Queensland, Australia, with Anthropic as its anchor tenant, was overwhelmingly opposed in public submissions. A filing lodged for developer Zerra DC counts 2 submissions in support, 1 neutral and 595 opposed or asking for a delay. Reuters says five objections came in before Anthropic was named as tenant on 15 September, and the rest between then and 11 November, when submissions closed. Objectors mostly cited water use, electricity demand and prices, noise, road damage and pressure on housing. The planned computing load is 1.44 GW. Zerra DC's spokesperson said the submissions 'reinforce that people want a greater say in how the project develops'; Anthropic declined to comment. Why it matters to the watch: this is local opposition over water and power to a 1.44 GW AI compute site, a constraint on where frontier-scale AI compute can be built.
On 20 November 2026 AZFamily/KOLD reported that a Maricopa County judicial officer dismissed a felony aggravated-assault charge in a Gilbert police case and questioned the weight of the evidence because officers used generative AI to help write the probable cause statement. The officer said there was "limited information provided about the extent of the usage and contribution of generative artificial intelligence". Gilbert police said they have used Axon Draft One for more than a year and that officers must review and verify every report. ABC15 reports that Commissioner Daniel Fenzel released a DUI suspect in a Department of Public Safety case, and that a DPS trooper's report disclosed it was written with Truleo AI. KOLD says the same judicial officer handled a second case involving a DUI suspect but does not name the officer. Why it matters to the watch: an AI tool drafted the document that justifies an arrest, and courts are now testing whether AI-written police records can carry evidential weight, which bears directly on AI-enabled surveillance of people.
JoyRide PH, a motorcycle-for-hire platform based in Manila, has confirmed that it was the target of a cybersecurity attack that compromised certain data within its app, Philstar reports. JoyRide says it has notified state regulators, brought in cybersecurity experts and added security controls. It asked customers never to share one-time passwords and not to respond to suspicious texts, and said its services are still safe to use. The extent of the breach is not yet known. Why it matters to the watch: this is a real cyberattack on a consumer platform, included under the owner's cyber-attack exception (18 November 2026); no AI link is reported.
On 20 November 2026 Amazon Threat Intelligence published a report that a suspected Chinese-nexus actor used autonomous AI agents running on Cairn against IP addresses associated with universities in East Asia outside mainland China, over an approximately 48-hour period beginning 8 September 2026. Amazon says it identified a publicly hosted Cairn instance on 10 September 2026. Amazon says the agents found a zero-day in an open-source application and used it to gain full control of a server, and that in about 11 minutes they went from mapping the application's API to arbitrarily writing files on the server. Amazon attributes the activity to a Chinese-nexus threat actor with moderate confidence. Why it matters to the watch: Amazon says the actor gave the agents the goal "obtain shell access" and that the agents conduct hacking operations with minimal human involvement, so this is an attack carried out by AI systems.
Americans for Responsible Innovation (ARI) released a report on 20 August 2026 calling for the federal government to designate AI as critical infrastructure. The American Prospect quotes that report as calling for an executive national security memorandum granting AI data centres special critical-infrastructure status, and instructing the intelligence community to 'conduct intelligence estimates on AI sector risks and … share those with industry'. The Prospect reports on 20 November 2026 that AI industry groups are pushing for the designation. It reports, citing two former counterterrorism officials who spoke on background, that it would increase surveillance by DHS and CISA and information sharing inside the FBI; the officials are unnamed. The Prospect also alleges that a June 2026 American Security Project report cited one of its articles for a homemade-explosives claim the article does not support; that report is offline. Why it matters to the watch: the designation would tie AI labs and data centres to federal intelligence-sharing and surveillance powers, which bears on AI governance and, per the Prospect, on government surveillance of critics of AI models.
On 20 November 2026 the Information Commissioner's Office (ICO) published a consensual data protection audit of the Metropolitan Police Service. It gave live facial recognition (LFR) and retrospective facial recognition (RFR) each a 'Reasonable' assurance rating. Recommendations: LFR five high, five medium and one low; RFR three urgent, 11 high and four medium. The executive summary says the data protection officer 'appears to be detached from the oversight and management of RFR', and that auditors 'could not gain assurance' that probe images are kept within retention periods. Operator Initiated Facial Recognition was not operational and no approved pilot documentation was available, so it was not assessed. Why it matters to the watch: the audit covers the Met's live and retrospective facial-recognition systems, and the regulator says that searches of the Police National Database after a failed match on the Met's Reveal system are 'not appropriately documented' and could run 'without appropriate oversight'.
Spain's Council of Ministers approved a Royal Decree on data-centre energy, environmental, resilience and digital-sovereignty requirements on 20 November 2026, according to the Ministry for the Ecological Transition (MITECO) release. For data centres above 10 MW, the decree requires 80% renewable consumption with hourly correlation, until renewables exceed 90% of the national electricity mix. The matching renewables must be connected at most 36 months before the data centre starts operating. Non-compliance can suspend grid-access and connection rights for up to three years. The methodology the ministry is developing foresees WUE at or below 0.1 and PUE at or below 1.25. Minister Sara Aagesen says that without the decree, electricity bills would rise by 5,000 million euros; the release presents this as the government's own scenario. Why it matters to the watch: it sets grid-access conditions for data centres above 10 MW, and data-centre power demand is the grid-load link to AI compute; the AI link is editorial, as the passages checked do not tie the rules to AI.
Taiwan has flagged nearly 2,700 social media accounts to Meta, mostly on Threads, which it says appear to be part of a Chinese information warfare campaign to influence the island's elections next week, according to three government officials, who spoke on condition of anonymity, and internal reports cited by Reuters. A second internal report cited by Reuters says 260 of the accounts shared more than 2,150 AI-made videos, which drew 2.35 million interactions and mainly targeted DPP politicians, including Puma Shen, the DPP's candidate for Taipei mayor, who denies the allegations. Meta said it removed the flagged accounts, many already disabled by its systems, and that none of these "low-quality fake accounts" had posted about election-related topics and most had no followers. Why it matters to the watch: the campaign reportedly used AI-made videos to target a democratic election, which is AI-enabled manipulation.
On 20 November 2026 TikTok's newsroom said it was 'starting to globally roll out Likeness Detection, an opt-in tool that helps eligible creators find and flag AI-generated content on TikTok that may use their likeness without their consent'. The same post says TikTok has also launched a transparency tool that lets anyone check whether content was made or edited with TikTok's own AI creation features. The post does not give eligibility criteria, detection accuracy, the countries that get the tool first, or removal timelines. Why it matters to the watch: it places part of the work of detecting AI-generated impersonation on the people being impersonated, and the transparency check covers only content made with TikTok's own AI tools, not outside generators.
MeioNews reports that iFood identified unauthorised access to some of its systems on 14 November 2026. The access involved registration data of customers, restaurant representatives, partners and delivery workers, and the exact fields vary by group; for some delivery workers, bank agency and account numbers were involved. In a statement on 19 November, iFood said it has found no evidence of disclosure or misuse of the information accessed, and that passwords, credit cards, other payment methods, balances, bank statements and biometric data were not accessed. Services continue to operate normally. iFood says it has notified Brazil's National Data Protection Authority (ANPD) and that the investigation is ongoing. Why it matters to the watch: a real unauthorised intrusion into a consumer platform's systems, included under the owner's cyber-attack exception (18 November 2026); no AI link is reported.
Expedia is investigating a cyberattack after a Russian extortion group, n0n, claimed to have stolen nearly 1 billion records. Expedia told Skift that it is aware of a supply chain attack, reported by security researchers, in which attackers compromised a third-party data platform provider. The group's claim of nearly one billion records and its attribution to n0n are unverified, and the sources read do not show Expedia confirming either. Why it matters to the watch: this is a reported supply-chain attack on a major travel platform, included under the owner's cyber-attack exception (18 November 2026); no AI link is reported.
161 with an event in the past 7 days, showing the 12 newest. 319 published signals in the current snapshot. 319 signals: 307 reviewed by editors, 12 published automatically after an AI check. 40 sources are monitored for new items; each is checked before it is published. This is a curated selection, not a complete record of global events. See all on Latest →
A global perspective.
Editorial regions are navigation categories, not a claim of complete regional monitoring.
Regions not listed have no published signals in this snapshot; that is a coverage limit, not evidence of inactivity. 131 published signals have no established country and are listed without a region.