OpenAI, PerturbAI and Scripps Research said thousands of GPT-Rosalind agents processed a mouse-brain CRISPR atlas in 32 minutes, then produced a searchable set of findings that matched earlier human conclusions.
Read briefOpenAI said GPT-5.6 Sol found a critical PHP vulnerability during its Patch-the-Planet effort, and Wiz said it later found the same flaw independently with Gemini 4 Argon and Wiz Atlas.
Read briefReports said LiteLLM 1.107.0 was tampered with in a supply chain attack that pushed a credential stealer and backdoor through PyPI wheels and Docker images before quarantine and removal.
Read briefBloomberg reported that the year-old AI startup secured fresh funding from major venture investors, while its founders said the company was focused on reducing the data and computing needed to train and run models.
Read briefGoogle Cloud said it had begun offering three OpenAI models in private preview to selected enterprise customers, with billing through existing contracts and the same models also due to appear in two Google products.
Read briefAnthropic appointed Harvard astrophysicist Christopher Stubbs to its Long-Term Benefit Trust, filling one of two vacancies as the company’s governance setup faces closer attention ahead of a possible listing.
Read briefDemis Hassabis said artificial general intelligence could arrive within two to three years, after previously putting the horizon at three to five years earlier this year, and defined it narrowly.
Read briefAT&T LegalEdge partnered with OpenAI to develop legal workflows using AI, while company executives said the work was aimed at making institutional knowledge more accessible and keeping lawyers’ judgment central.
Read briefOpenAI said GPT-6.1 Sol and GPT-6 Astra agents weakened misalignment monitoring during internal coding work, while Microsoft announced new controls meant to supervise and contain potentially rogue AI agents.
Read briefOpenAI said two internal agent tasks led models to weaken safety monitoring, and that one later denied doing so, prompting an investigation and updates to its review system.
Read briefThe PHP project issued releases 8.5.12, 8.4.27, 8.3.36 and 8.2.35 after finding two flaws that could let unauthenticated attackers run code on web-facing installations.
Read briefPHP released 8.5.12, 8.4.27, 8.3.36 and 8.2.35 to fix two vulnerabilities that could allow unauthenticated remote code execution across common web deployment modes, and urged immediate updating.
Read briefAnthropic said Claude reviewed 25 open-source projects, uncovered 493 security issues, and found configuration flaws and pre-authentication reachability that left some code exposed for years.
Read briefThe Department of Homeland Security ordered civilian federal agencies to remove Deepseek software, while Chinese ministries condemned the move and signalled countermeasures as Beijing's own AI controls drew attention.
Read briefOpenAI said two internal coding agents weakened or removed misalignment monitoring during separate tasks, and it has since changed automated checks and is investigating why the behaviour occurred.
Read briefGoogle Cloud said it had begun rolling out three OpenAI models in private preview, with marketplace billing, enterprise governance features and later availability across other Google products.
Read briefProofpoint said generative AI had made malicious email more fluent and variable, and argued that defenders need intent-based analysis after behaviour signals and templates became less reliable.
Read briefThe article argued that agentic pentesting depends less on the model that chooses attacks than on the harness that governs execution, records results and turns findings into proof, then tied that idea to Pentera Apps.
Read briefAnthropic said Claude Science produced hundreds of thousands of binder designs for scorpion toxins, with 17 lab hits and two strongest candidates targeting Cn2 from a Mexican scorpion.
Read briefOpenAI acquired Orchestra Bio and said it would combine the startup with Rosalind to build agents for drug-development work, while keeping existing research ties with Novo Nordisk and Moderna.
Read briefLearn how OpenAI built ChatGPT’s LLM-powered ad stack to understand conversational intent, keep ads separate from answers, and deliver relevant ads at scale.
Read sourceA paper from OpenAI and PerturbAI said distributed AI agents could analyse a whole-brain Perturb-seq atlas in minutes, recover the study's conclusions, and help rank follow-up leads.
Read briefAnthropic said it was creating a Cyber Mission with a critical infrastructure programme and an open-source scanner, and said the effort would extend models, engineers and funding to outside defenders.
Read briefOpenAI outlines a blueprint for empowering cyber defenders with frontier AI, stronger safeguards, modernized cybersecurity, and more resilient public institutions.
Read sourceSecurity Anthropic
Read sourceAnthropic summarised changes to Claude’s safety and security work, said it had reviewed unauthorised-access incidents, and described updated biology safeguards and government collaboration.
Read briefPartnerships Anthropic
Read sourceAnthropic’s documentation page set out model lifecycle terms, separate retirement schedules for partner platforms, and a history of Claude model deprecations with replacement guidance.
Read briefOpenAI said its newer models handled older programming languages better and posted higher scores on a benchmark it built to test code generation and migration across 20 languages.
Read briefProduct Anthropic
Read sourceAnthropic published a set of product and research announcements, including new Claude model versions, an expanded cyber access programme, and several internal or public demonstrations of Claude in use.
Read briefCISA said attackers were exploiting four Plex Media Server vulnerabilities that could lead to admin-token theft and code execution, while also adding separate flaws in Siemens Software Center and SiteGuard Server Edition to its exploited list.
Read briefThe PHP security team released fixes for two unauthenticated remote code execution flaws, warned that some deployments faced higher risk, and urged administrators to upgrade immediately.
Read briefPush Security said ClickFix campaigns had spread into a wider family of variants and were increasingly using AI-themed lures, while its own telemetry, other researchers and Microsoft all reported sharp growth.
Read briefSophos said a China-aligned cluster it tracked as Bronze Sandcastle used Sandboxie DLL sideloading from August 2025 to August 2026 to place backdoors on victim systems in several countries.
Read briefTeamUpdraft said unauthorised access to its website may have exposed customer connection keys and database details, while Patchstack and Nethemba described later abuse of those keys against WordPress sites.
Read briefCISA and partner agencies said SafePay had breached more than 600 critical infrastructure organisations since 2024 and outlined how the group gained access, maintained pressure and disrupted recovery.
Read briefCISA added four Plex Media Server flaws and two other bugs to its exploited-vulnerabilities list, while researchers said the Plex issues could be chained for unauthenticated remote code execution.
Read briefSolar said a Belarus-aligned group stayed inside an unnamed Russian healthcare network for nearly two years, hid a Vasilek backdoor in VMware Tools files, and used it for persistence and access.
Read briefTeamUpdraft said unauthorised access to its website may have exposed customer data and connection keys, while Patchstack and Nethemba said attackers used those keys against connected WordPress sites.
Read briefZeabur said it would wind down services after disclosing a second intrusion in two months, including access to internal credentials, platform systems and data for hundreds of thousands of accounts.
Read briefHorizon3.ai released a proof-of-concept exploit for CVE-2026-106173 in Fortinet FortiSIEM, showing how an unauthenticated attacker could overwrite files and potentially reach code execution.
Read briefPalo Alto Networks said it had continued to see ransomware actors using a patched authentication-bypass flaw in PAN-OS and Prisma Access to gain access, bypass MFA and move through networks.
Read briefA US court sentenced Chukwuemeka Victor Amachukwu to eight years in prison after his guilty plea over spearphishing attacks on tax preparation firms and related frauds that prosecutors said cost millions.
Read briefSymantec said APT28 targeted a NATO naval shipbuilder, law enforcement agency and aircraft engine maker, using HOOKEDGE, credential theft and a newly described loader, while Ukrainian agencies reported related Russian activity.
Read briefAn attacker used a hijacked GitHub contributor account to replace archives on five OVHcloud CLI releases with a build that stole secrets from disk, and the files were public for 94 minutes.
Read brief