Chatbot’s Fake Intel Nearly Sent Jets After a Ship as AI Labs Get Sued

AI & Machine Learning

Military aircraft were already airborne when US officials realized the intelligence behind an armed operation against a Chinese vessel had been fabricated by an AI chatbot, and the operation was aborted at the last minute. According to reporting first published by CNN, an analyst at Special Operations Command used an unspecified chatbot to combine open-source data with classified signals intelligence. The model misread the ship’s cargo manifest and claimed it carried nuclear weapons program components, and it was then used again to format the false finding into an official-looking summary. The episode came during the war with Iran and could have set off a confrontation with China. GovAI’s Jake Steckler argued for adding safeguards and training service members on LLM uncertainty rather than abandoning AI, and no formal government response has been reported. Source: TechCrunch Verified: True

Anthropic is operating a wet biology lab in the Bay Area where its AI models can run physical experiments and test their own theories against real-world results. Eric Kauderer-Abrams, the company’s head of life sciences, said the work targets fundamental biology rather than drug discovery, and that lab results remain the final test of any AI-generated hypothesis. Anthropic bought the biotech startup Coefficient Bio in April 2026 and also runs a Life Sciences Verification Program that gives vetted researchers access to its most powerful models. The disclosure lands days after CEO Dario Amodei called for an industry slowdown and named bioterrorism as a leading AI risk, which sharpens the tension between AI-driven science and biosecurity. Source: TechCrunch Verified: True

Consumer Hardware

Google began rolling out Home MCP on September 16, letting outside AI agents such as Google Antigravity, Claude, Hermes and OpenClaw read and control a Google Home setup. Agents can analyze camera footage across rooms, review device state history, send voice messages through Google Home speakers and build custom dashboards, across Nest cameras, doorbells, thermostats and Matter bulbs. Google says it applies rate limits and blocks sensitive actions such as unlocking doors. It also warns that connecting an agent can cause unexpected or undesired behavior. The rollout starts with Google Home Premium Advanced users in US English and requires creating a Google Cloud project to grant permissions. Source: 9to5Google Verified: True

“Talking Tilly,” a video-calling service featuring the AI-generated character Tilly Norwood, is shutting down permanently on September 27 after drawing scrutiny over its data practices. The character is powered by Google’s Gemini through the Tavus platform, and its operator, UK-based Xicoia, calls it an awareness project showing how far AI video has come. Callers must pass a facial age check through the Spanish firm Didit, and the service analyzes their mood in real time from camera and voice. Recordings and transcripts are processed by US providers and kept for up to eight weeks, and the company relies on “legitimate interests” rather than explicit consent. One caller’s recording was wrongly flagged for abusive language that never occurred, and UK compliance means the face scanning applies worldwide. Source: BleepingComputer Verified: True

Cybersecurity

Cisco warned that a maximum-severity authentication bypass in Identity Services Engine, tracked as CVE-2026-76460, is being exploited in attacks. The flaw sits in the ISE API and lets a remote attacker bypass the web management interface on any configuration, with no workaround. Fixes ship in ISE 3.1 Patch 12, 3.2 Patch 11, 3.3 Patch 12, 3.4 Patch 7 and 3.5 Patch 4. CISA gave federal agencies three days to patch. Because ISE controls network access policy, a compromised appliance can hand attackers a foothold across the whole network. Source: BleepingComputer Verified: True

A flaw dubbed Plugin4Shell lets an attacker who controls a plugin repository swap the code an AI coding agent installs, even when the plugin version is pinned. Air Security, which reported it in June, found that agents fetch pinned versions by commit hash without checking that the code matches, so a branch named like a hash can redirect the install. Claude Code fixed it in version 2.1.179 and OpenAI Codex in 0.146.0. GitHub Copilot has no patch, and Google will not fix Gemini CLI and recommends moving to Antigravity. Plugins installed from GitHub are not exposed, because GitHub blocks branch names that resemble commit hashes. Source: The Hacker News Verified: True

Three researchers at the AI-assisted security firm Hacktron chained two flaws to take over the ChatGPT and Codex accounts of multiple OpenAI employees, and The Hacker News credits Claude Opus 5 in its headline. The first was an out-of-bounds read in the libheif image library (CVE-2026-32882) used by OpenAI’s Discourse help forum, which gave them code execution on the forum server. The second was the shared “Sign in with OpenAI” login between that public forum and internal staff tools, which let them hijack staff sessions and reach an internal GitHub repository. They opened one harmless pull request and say they read no source code and touched no customer data. OpenAI fixed the login flaw about 14 hours after notification and paid a $6,500 bounty. Source: The Hacker News Verified: True

Image-sharing service Gyazo disclosed that attackers exploited a server vulnerability on September 11 and stole about 23.6 million user records. The haul includes names, email addresses, password hashes, session tokens, Google single sign-on emails, subscription details and billing information. Roughly 490 million image metadata records were also exposed, containing upload IP addresses, user-agent strings, EXIF location data, OCR text and image titles. Gyazo detected the intrusion on September 12 and patched the flaw, took the platform offline for maintenance, and is notifying users and advising password changes. It has also contacted authorities. Source: BleepingComputer Verified: True

Enterprise Infrastructure

Nvidia introduced power-management software under its DSX platform that squeezes more GPU capacity out of data centers limited by the electric grid. Typical facilities run near 80% of capacity, which strands about 20 MW in a 100 MW site. DSX MaxLPS links compute with power cabinets and chillers to adjust operations in real time, and DSX Flex lets operators pause non-critical workloads during grid demand spikes. In a Lambda demonstration, 19 nodes fit into space normally holding 16, lifting cluster throughput and performance per watt by about 24%. The software mainly benefits Nvidia GPUs, and Nvidia’s Dion Harris framed the goal as more performance per gigawatt. Source: The Register Verified: True

Fujitsu said it will sell its custom Monaka Arm server processor to cloud operators and server makers starting in November 2026. The Armv9 chip has 144 cores across four 2nm compute dies with 5nm SRAM chiplets, plus 12 DDR5 channels and PCIe 6.0. Fujitsu offers 1U and 2U servers, including a liquid-cooled multi-node design for HPC, and says the chip handles double the inferencing output of rival CPUs. It rates the servers for 40°C ambient air, or 45°C with liquid cooling. Fujitsu is pitching Monaka as a “sovereign” option, built in Japan for European and other buyers wary of relying on foreign infrastructure. Source: The Register Verified: True

Policy & Regulation

A private antitrust class action filed Friday in federal court in San Francisco accuses Anthropic, OpenAI, SpaceXAI and Google of illegally agreeing to slow AI development. Four paying subscribers to ChatGPT, Claude, Grok or Gemini brought the suit and want class status for all paid subscribers, arguing the coordination reduced the value consumers get. It cites Dario Amodei’s September 12 essay urging an industry slowdown, the public agreement from rival chiefs the same day, and a July statement by lab employees acknowledging competitive pressure against slowing alone. The suit turns the lab safety talks reported earlier this week into a formal antitrust claim. The companies had not responded to requests for comment as of Saturday. Source: Local10 Verified: True

Virginia Governor Abigail Spanberger signed an executive order tightening how data centers are approved in the state that hosts more of them than anywhere else. It bans nondisclosure agreements tied to data center projects and removes large facilities from expedited permitting, requiring local approval for anything above 25 megawatts. It also restricts water-hungry cooling towers in water-scarce areas and encourages retrofitting backup generators or moving to batteries or fuel cells. Loudoun County paused certain applications for up to 12 months at the same time. The order follows New York’s July halt on projects above 50 megawatts and Pennsylvania’s August limit on permits without binding commitments. Source: The Register Verified: True

President Trump said on Truth Social that he will create an “AI Force,” modeled on the Space Force from his first term, and appoint an AI czar. The post said “Only High I.Q. individuals need apply” but gave no detail on what the new body would do or how it would fit with existing agencies. Trump also polled followers on renaming artificial intelligence, offering Superior, Extreme or Supreme Intelligence. He called AI safety concerns a Democratic hoax, which sits at odds with the state-level safety rules and lab-led slowdown talks taking shape elsewhere. Source: TechCrunch Verified: True