Anthropic said it had started giving METR employee-like access so the nonprofit could assess its models and safety practices, while it also described other evaluator partnerships and a three-tier access system.
Read briefMeta AI Research released Muse Spark Open, an open-weight multimodal model, and said it could run locally on two Apple M5 Ultra Mac Studios using the MLX framework.
Read briefAnthropic said it had begun giving METR employee-like access to its systems, while setting out a tiered model for other independent evaluators and limiting what they could see or publish.
Read briefGoogle presented Gemini 4 Argon as a single robotics model for control, simulation and training-data review, but its examples and benchmark claims also showed heavy prompting and mixed results.
Read briefBank of America said 1,000 contractors and 3,000 employees were working on software vulnerabilities identified by AI models, and Brian Moynihan said the bank spent a few hundred million dollars on the work and anti-money-laundering requirements.
Read briefAnthropic told investors third-quarter revenue reached $16.8 billion and adjusted profitability improved, but slower growth, distillation revenue and softer IPO sentiment left its November listing plans under pressure.
Read briefPrevidian said attackers were using an authenticated remote-code-execution flaw in Langflow, while honeypot data showed more than 1,400 attempts against its sensors since 19 September 2026 and signs of credential abuse.
Read briefHaydn Belfield said Google DeepMind had not seen measurable AI-driven acceleration in research, even as automation expanded in coding and internal benchmarks showed limits on harder tasks.
Read briefPreparing for the 2026 US midterms Anthropic
Read sourceThe Information, citing employees and other people familiar with the work, said OpenAI expected to release one or two models by year-end, while its main upgraded model still needed alignment and testing work.
Read briefAnthropic said smaller AI models may soon self-replicate on Internet of Things hardware, while its tests showed the idea remains difficult to realise and would be hard to shut down.
Read briefSee how Amp pushes the frontier of what agents can do with GPT-6 Astra, Ultrafast, GPT-Live-1, and Sign in with ChatGPT.
Read sourceSentinelOne said it had expanded Wayfinder Frontier AI Services with OpenAI Daybreak models, plus code-risk and compromise-assessment functions, starting with GPT-5.6-Cyber and rolling the additions out in private preview.
Read briefXBOW said it combined three Chrome vulnerabilities into a Windows code-execution chain in six days using AI agents built on OpenAI’s public GPT-5.6 Sol model, and Google paid the company $250,000.
Read briefAnthropic opened a Singapore office and signed a memorandum of understanding with the Singapore government on AI cooperation, with both sides focusing on safety, security, adoption and wider impacts.
Read briefPackages uploaded under names resembling yay and paru were found in the Arch User Repository with an install script that fetched remote code as root, and the AUR team removed them.
Read briefCisco Talos and Recorded Future Insikt Group said the threat actor Silent Lynx ran four intrusion chains from June 2025 to August 2026 against government targets across the CIS, including one using CVE-2025-8088.
Read briefANY.RUN said it found 1,258 public sandbox analyses tied to campaigns impersonating AI services between April and October 2026, and argued that MSSPs could use its tools to spot them sooner.
Read briefPalo Alto Networks released fixes for a PAN-OS management-interface flaw after limited exploitation, while Cisco separately warned of an exploited Secure Email Gateway bug and CISA added both to KEV.
Read briefAhsay Backup issued new hotfixes for AhsayCBS after saying earlier August updates had not fully fixed two flaws that Huntress said were being chained for remote code execution and payload deployment.
Read briefPalo Alto Networks released updates for a PAN-OS remote-code-execution flaw after saying it had seen limited exploitation, while CISA ordered federal agencies to mitigate the issue by 17 October.
Read briefCisco said attackers were exploiting a Secure Email Gateway vulnerability that could let an unauthenticated sender write to the appliance and, if successful, run commands as root.
Read briefCisco said it had fixed CVE-2026-76519 in Secure Email Gateway, a flaw in its Advanced Malware Protection feature that it said had been actively exploited and could let attackers run commands as root.
Read briefMicrosoft, Health-ISAC and Japan's JC3 said they disrupted three phishing kits after an investigation that began in late 2025 and linked them to a large phishing operation.
Read briefMicrosoft said an investigation with Japan’s National Police Agency uncovered a Chinese-language phishing marketplace tied to several kits, then led to court-backed seizures, account deactivations and domain blocking.
Read briefCISA told organisations using Ivanti Connect Secure to install the latest release after disclosure of an authentication-bypass flaw, while Previdian said honeypots recorded 105 exploitation attempts from 50 IP addresses.
Read briefThreatFabric said the Android banking trojan Hirudo used dropper apps and fake update prompts to install bank-specific replacements that bypassed app checks and let operators steal credentials and control devices.
Read brief