OpenAI, Anthropic, Meta and Microsoft said they were stepping up election monitoring and content controls as the Brennan Center said some AI image tools still handled fake-election prompts unevenly.
Read briefStanford researchers tested 50 widely cited claims about large language models against current systems and found that many had weakened or expired, with results varying by model and claim type.
Read briefGoogle Threat Intelligence Group and Mandiant said LLM hijacking campaigns were producing fraudulent cloud AI bills, with some victims reaching $1 million through stolen keys and abused infrastructure.
Read briefChainstack said an intruder reached part of its infrastructure through a third-party component, accessed customer account data and password hashes, and may also have exposed node credentials.
Read briefGerrit De Vynck / Washington Post : Sources: UK-based Google DeepMind staff urge leaders to drop non-disparagement clauses barring public warnings about AI risks even after they leave the company — Employment contracts deter staff from warning the public about AI risks, current…
Read sourceGoogle said attackers had been hijacking enterprise LLM access through stolen or exposed credentials, with some incidents producing six-figure daily charges and a separate relay service rerouting coding tools through third-party servers.
Read briefOpenAI appointed former Army Secretary Dan Driscoll to a new national security post as it expanded its defence work and positioned itself for further government business.
Read briefThe Information : Sources: Anthropic has told investors in recent meetings that its 15-person chip design team is using Claude for engineering and design work on its custom chips — Anthropic's AI models have helped uncover serious cyber vulnerabilities and transform the…
Read sourceAnthropic launched Message Threads as a public beta on Claude, saying it cuts the data sent between turns and reduces time to first token for very long conversations.
Read briefOpenAI said it had raised rewards for vulnerability reports in products such as ChatGPT and Codex, added loyalty and campaign bonuses, and planned a security sprint with Cathedral.
Read briefAnthropic said Claude Code now blocked attempts to copy existing video games, after users had been using the tool and other AI systems to produce unofficial ports and knockoffs.
Read briefHeimstaden said it chose OpenAI in June 2025 and has used ChatGPT Enterprise and Codex across letting, data analysis and procurement while keeping employees responsible for review and approval.
Read briefGoldman Sachs restricted advanced AI coding tools to engineers after previously letting traders and bankers use them, with the bank citing productivity and risk management for the change.
Read briefOpenAI updated its security bug bounty programme with higher base awards, report-quality multipliers, loyalty bonuses and campaign pools, taking the maximum payout for some findings to $172,000.
Read briefOpenAI is preparing its next internal security sprint with Cathedral Systems to test and strengthen its defenses using AI.
Read sourceOpenAI said its Pando system was built to handle more commits from developers and agents while keeping Git compatibility, reducing push latency and changing how its servers store and serve repository data.
Read briefOpenAI said its new non-binding programme would let future data centre campuses reduce electricity demand during grid stress, while it works with utilities and regulators on local arrangements.
Read briefThe newsletter summarised a real-time benchmark for AI systems, a Defence Intelligence Agency report on space conflict, a reasoning-architecture paper from Metis, and Google DeepMind’s Genie 4 preview.
Read briefBinance said a June supply-chain compromise and a September fake recruitment scheme led to a wider intrusion that ended in unauthorised wallet transactions and a rebuild of affected systems.
Read briefBarracuda said two vulnerabilities in its products could be chained to bypass authorisation controls and gain root access on affected devices if an attacker reached the administrative interface.
Read briefBarracuda Networks told customers to patch two vulnerabilities in its Email Security Gateway appliances after saying combined exploitation could let an attacker gain root access.
Read briefBitdefender said a 3AM affiliate attacked a UK school estate in September, used stolen credentials and domain trusts to spread, and encrypted about 37 machines while researchers examined possible AI-assisted command patterns.
Read briefGoogle said researchers can now earn $250,000 for Root-in-Cloud flaws that expose full administrative control, while also tightening rules around testing and automatic review of some production-infrastructure exploits.
Read briefRed Hat shipped fixes for a Linux kernel IPv4 TCP use-after-free that could permit remote code execution, while Canonical kept Ubuntu updates for three LTS releases in its testing channel.
Read briefESET said the Russia-aligned Gorbag group had targeted Ukrainian organisations with stealers and a reverse shell, while Trend Micro linked earlier June activity to a patched WinRAR vulnerability.
Read briefHackers accessed Harvard's alumni and fundraising systems in November 2025, and Have I Been Pwned later said leaked material exposed data linked to 851,000 accounts, including contact and profile details.
Read briefSynology said it had seen threat actors abusing a flaw patched in December 2024 against outdated DiskStation Manager versions, after customers reported C3RB3R ransomware on Synology devices.
Read briefVulnCheck said it saw exploitation of CVE-2026-58115 against honeypots using Siemens SIMATIC IoT2050 Advanced devices, while Siemens had already issued a fix and mitigation steps for affected Node-RED installations.
Read briefDefused Cyber said attackers were trying to exploit a critical flaw in the CryptoPayment Gateway WordPress plugin, which NIST said could let unauthenticated users delete files and recover wallet credentials.
Read briefGoogle said it and partners disrupted the XCPE botnet, linked it to Videostrong Technology Co, and cut off infrastructure tied to additional malware families and compromised devices.
Read briefThe recap assembled reports on a Binance theft attributed to TraderTraitor, covert infrastructure used by Volt Typhoon, a Polymarket suspension after a third-party compromise, and several other incidents.
Read briefDream Research Labs said DarkBlinders targeted Israel and the Kurdistan Region of Iraq between August and October 2026, while Semperis said Prince of Persia ran a server from the United States this year.
Read briefIBM said two critical Langflow vulnerabilities could let unauthenticated attackers bypass login or run code, and it said version 1.13.0 was the only fix for releases 1.0.0 through 1.12.8.
Read briefPalo Alto Networks Unit 42 said a campaign targeting government-linked Linux systems used AI-generated scripts, web shells and a custom Go trojan, while related research described abuse of agentic penetration-testing frameworks.
Read brief