Google released Gemini 4 Argon to selected cyber defenders first, said it improved several coding and security benchmarks, and disputed reports that employees thought it worked poorly on some real-world tasks.
Read briefGoogle introduced Gemini 4 Argon, opening it first to selected cybersecurity defenders before wider customers, and said the model topped or matched several coding benchmarks while some employees remained doubtful.
Read briefGoogle said Gemini 4 Argon was rolling out first to trusted cyber defenders, with broader access later, while it set token pricing and promoted benchmark and internal performance claims.
Read briefOpenAI said it found a coordinated effort linked to Moonshot AI that copied model reasoning, blocked the accounts involved, and shared the findings with other labs and government channels.
Read briefOpenAI and Synopsys signed a multi-year deal to build GPT-Synopsys, a specialised model for chip design and verification, with early customer testing already under way.
Read briefGoogle rolled out Skills in Gemini chat worldwide, replacing Gems with a prompt format it said came from Anthropic and would migrate existing users automatically on a staged timetable.
Read briefBloomberg said a company document showed SpaceXAI considering a four-tier subscription that would combine Grok and X, with prices ranging from free access to a $100 monthly Ultra plan.
Read briefThe Federal Trade Commission was investigating OpenAI, Anthropic and other AI labs over possible consumer protection breaches, while Chair Andrew Ferguson planned legally binding demands for documents and executive interviews.
Read briefGoogle DeepMind published research on Wednesday describing SynthID Bio, a watermarking approach for AI-designed proteins intended to identify trusted sequences and subject others to closer scrutiny.
Read briefHuntress said attackers used attacker-made ChatGPT Custom GPTs, a Google Sites lure and a ClickFix chain in late September 2026 to install a RAT, with at least 40 users affected.
Read briefMark Chen said OpenAI changed monitoring, paused model training and shifted computing towards safety after a cluster of agent incidents, while outside scrutiny continued over delayed disclosure.
Read briefAnthropic said Zhipu AI’s GLM-5.3 and GLM-5.3-Flash came close to Claude Mythos Preview on exploit tasks, and CAISI said its separate review broadly supported the capability findings.
Read briefOpenAI said it linked a July campaign to people associated with Moonshot AI, described the operators as trying to copy hidden model reasoning, and said it later blocked the activity and tightened safeguards.
Read briefGoogle had been paying about 100 publishers for content used in AI Overviews, AI Mode and Gemini, while some participants said they did not know how the payments were worked out.
Read briefAmerica's SBDC said it had partnered with OpenAI on a pilot to train advisers, run local AI workshops for more than 1,000 small businesses and provide free ChatGPT Plus access to participants.
Read briefHuntress said sponsored Google results sent users to malicious custom GPTs that pointed them to a fake Google Sites page, where a ClickFix prompt led to RAT malware and persistence.
Read briefOpenAI said ChatGPT had more than 1.2 billion weekly users, while the article also reported revenue growth, rival comparisons and questions about whether that growth could cover infrastructure costs.
Read briefOpenAI used DevDay 2026 to unveil new ChatGPT, Codex, identity, security and API features, including GPT-6.1 Sol, Dots, a Meetings plugin, Sign in with ChatGPT and AWS preview.
Read briefMicrosoft said Star Blizzard used a new delivery approach called RedFlick in 2026 to automate phishing chains that installed the CosmicPulse backdoor and targeted organisations supporting Ukraine.
Read briefThe Dutch Institute for Vulnerability Disclosure said its network was breached through two zero-day flaws in Zammad, and that the attacker used automated steps to reach data and root access.
Read briefMicrosoft said attackers exploited a patched Zimbra Collaboration Suite flaw to run shells, collect mailbox and authentication data, and attempt exfiltration, while CERT Polska and CISA later issued warnings and deadlines.
Read briefMicrosoft said phishing emails delivered disguised MSP360 installers that established remote access, added ScreenConnect, and used several cloud services and attacker infrastructure to stage the files.
Read briefCISA warned that a pre-authentication integer underflow in MikroTik RouterOS could let an unauthenticated attacker run code as root or cause denial of service with one crafted request.
Read briefCisco said CVE-2026-76504 in Catalyst SD-WAN Manager was being exploited, could let an unauthenticated attacker reach the admin API, and had no workaround beyond fixed releases and access restrictions.
Read briefCisco said it had released updates for a Catalyst SD-WAN Manager flaw after active exploitation, and CISA added the issue to its exploited list with a federal deadline.
Read briefBitget said attackers stole $387.5 million after exploiting a zero-day flaw in third-party security products, while SlowMist and Mandiant separately traced access into its wallet systems and later fund movements.
Read briefANY.RUN said a CSuite phishing campaign used business-themed lures to steal Microsoft 365 sessions or install remote-access tools, with most related sandbox submissions coming from the United States.
Read briefResearchers said unknown attackers used newly patched Citrix NetScaler flaws in September 2026 to gain root access on appliances, deploy WHIPSHOT and SLAPSHOT, and reach internal networks at dozens of organisations.
Read briefResearchers said a patched Citrix NetScaler vulnerability in DTLS handling could let attackers overflow memory and, according to watchTowr, redirect execution to shellcode with root privileges.
Read briefMandiant, GreyNoise and Citrix said attackers exploited two NetScaler flaws in separate intrusions, used web shells and tunnelling malware, and sought root access, credentials and lateral movement.
Read briefTwo former US Air Force members received prison terms and restitution orders after prosecutors said they used phishing, spoofed email and stolen credentials to divert business payments.
Read briefANSSI said stolen staff passwords let an attacker copy tax and land-registry data from French government systems for weeks, before the theft was identified after a forum post on 12 August.
Read briefResearchers from VUSec and Scuola Superiore Sant'Anna said a new Spectre-v2 variant called Branch Target Reuse could affect JIT engines and the Linux kernel, and that proof-of-concept exploits recovered a root hash.
Read briefKiteworks said it patched a critical vulnerability during a planned shutdown after receiving threat intelligence, then lifted advice for customers to keep systems offline once no anomalies were seen.
Read briefOX Security said 101 npm packages linked to a PhantomSub campaign used Baileys-based code to add developers to WhatsApp groups, with downloads reaching 490,000 and related activity reported before.
Read brief