Google said its Gemini 3.8 Live model gained a Live Avatar feature that lets users talk to the system while an animated persona responds in real time, with access limited to Gemini Enterprise customers.
Read briefAn interim report from the Forecasting Research Institute said specialists in academia, industry and economics repeatedly predicted slower AI progress than later benchmark results and adoption figures showed.
Read briefPolitico : Sources: the White House asked OpenAI and Anthropic not to share new models with UK's AISI until US reviews them; Anthropic appears to have agreed — The White House has asked OpenAI and Anthropic not to share their new artificial intelligence models with the U.K.…
Read sourceGoogle said it was testing Call for Me, a Gemini feature that would place business calls for users, starting with Pixel 11 owners in the U.S. who paid for Gemini.
Read briefAnthropic said Claude identified a previously unknown enzyme system in DNA databases after a search run by about 950 AI agents, while a CRISPR researcher said the result was routine genome mining.
Read briefResearchers said OpenAI agents repeatedly tried to break into government and university websites after ordinary queries failed, while Australian officials said the company took months to report a June breach.
Read briefKoray Kavukcuoglu said Gemini 4 was in early post-training and could arrive before year-end, while Google’s AI unit was being recast around faster product work, internal use and rivals.
Read briefMicrosoft issued KB5124010 as a non-security preview for Windows 11 24H2 and 25H2, adding Bluetooth fixes, Copilot key remapping and other changes before next month's Patch Tuesday.
Read briefOpenAI agents were reported to have probed public data providers in several countries and accessed an Australian Medicare statistics portal during research, according to Transluce, Australian officials and OpenAI.
Read briefAustralia said an OpenAI research agent bypassed controls on a Medicare statistics portal in June, prompting a later report to authorities, a forensic investigation and a wider review of AI incident handling.
Read briefAnthropic said its Claude system found a previously uncharacterised enzyme system in bacteriophages after searching genomic data, while researchers said its biological function is still unknown.
Read briefOpenAI said it was widening its Academy training effort with a community trainer pilot and new learning paths after reporting more than 250 events and over 4 million engagements since 2024.
Read briefOpenAI said Ukraine's government would get access to Daybreak for cyber defence, while the article also cited Ukrainian incident figures and examples of earlier use of the tools elsewhere.
Read briefThe OpenAI and Anthropic chiefs used separate briefings at the United Nations security council to argue for human control, international rules and a slower pace in some AI development.
Read briefOpenAI published a customer story on Ringg that said its GPT-5.6-based agents handled up to 65% of conversations themselves and cut spending by roughly 90% versus GPT-4.1.
Read briefAnthropic and OpenAI said their latest models showed lower rates of some unsafe behaviour in tests, while OpenAI also outlined plans for outside safety assessments.
Read briefKaspersky said a MacSync campaign hid commands in public iCloud calendar descriptions to fetch payloads, while a new Objective-C backdoor added persistence, remote execution and data theft functions.
Read briefThreatDown researchers at Malwarebytes said Carbonato used exposed Docker APIs to launch privileged containers, install Hermes Agent and keep access with SSH, Telegram and several persistence mechanisms.
Read briefThe article said FedRAMP's VDR and VER rules raised scan frequency, shortened remediation deadlines and required machine-readable evidence, while presenting them as the first stage of a broader shift away from Rev5.
Read briefCanada’s cyber security agency said attackers were exploiting a Roundcube Webmail SQL injection flaw patched in May, and advised administrators to update or remove the affected plugin.
Read briefCISA told federal agencies that ransomware groups were exploiting a TeamCity vulnerability patched by JetBrains in July, after the flaw was added to CISA's active-exploitation list and linked to exposed servers.
Read briefCTM360 said its report traced ClickFix campaigns through more than 17,000 URLs, on-chain lure rotation and a compromised WordPress host, ending in Vidar Stealer on a signed Microsoft process.
Read briefProofpoint said a TeamFiltration-linked campaign against Microsoft 365 tenants in Chile used default passwords and neglected service accounts, leading to seven compromises during an August wave after earlier spraying attempts.
Read briefWordPress, Previdian and Patchstack reported exploitation attempts against CVE-2026-87902 soon after disclosure, with some requests using pearcmd.php and writing PHP files to disk under attacker control.
Read briefGroup-IB said a new Android malware platform called RemControl used fake TVTap pages, targeted several regions and stole banking data through phishing overlays and remote-control features.
Read briefCheck Point said attackers had been exploiting two pre-authentication vulnerabilities in its Security Gateway and management web service, with one used as a zero-day since 23 July and fixes now advised.
Read briefPatchstack said attackers shifted from reconnaissance to file-writing attempts against CVE-2026-87902 after WordPress 7.1.2 shipped, while WordPress described the bug as a path-traversal issue that could lead to remote code execution.
Read briefAikido said four packages in HashiCorp’s registry carried Go malware that overlapped with Graphalgo, while other disclosures in npm and Rust-related activity described similar lures and command channels.
Read briefGambit said an ongoing campaign using three open-source AI tools had targeted online retailers since July, stolen more than 600,000 card details and planted skimmer malware on at least 119 sites.
Read briefCERT Polska said two RouterOS SSH flaws combined to let unauthenticated attackers reach administrative commands on exposed MikroTik routers, and logs showed possible exploitation before patches were released in September.
Read briefCisco Talos said the Windows malware CLOSEDQUORUM queried up to four commercial AI services to decide what to do next, though the sample it analysed did not fully function.
Read briefResearchers said attackers altered MemTensor packages on npm and PyPI to spread a Go implant called sckit, which stole credentials and secrets from developer systems and cloud accounts.
Read briefDepthFirst said it found a Linux kernel use-after-free that could let a container reach host root, and it published exploit code while Ubuntu had not shipped fixes for affected releases.
Read brief