Two employees from METR and Redwood Research worked on premises at OpenAI for three weeks this summer in an embedded assessment pilot. This team evaluated the systems OpenAI was using to monitor for misalignment from internal agents. A brief write-up of their work is now public.
#Safety / alignment#Agents#OpenAI
Uber and OpenAI announced an expanded partnership under which all of Uber's roughly 30,000 employees worldwide will use ChatGPT Work. The product, launched in July this year, is positioned as an enterprise-facing AI agent; Uber had previously piloted it with 1,000 employees across departments including finance, operations and product, where about 80% of early users were not engineers, using it to analyze operational data, localize training materials and build interactive dashboards. Uber did not disclose the cost of the rollout, saying it will track how employees use the tool and their output to guide future AI deployments.
#Industry news#Agents#OpenAI
Anthropic is about to list on NASDAQ, and its equity structure shows that no single VC firm holds more than 2%; Iconiq, Menlo and Lightspeed will be the biggest VC winners in this IPO. Reports say that for years, multiple VC firms passed on investing due to concerns over its special corporate structure and its insistence on prioritizing safety over commercialization, and Anthropic had also communicated to investors its goal of keeping any single investor's stake below 2%.
#Industry news#OpenAI#Anthropic
Two critical vulnerabilities have been found in PHP, rated 9.2 out of 10 in severity, affecting all versions since PHP 5 and enabling remote code execution. Sites including Wikipedia, WordPress, 4chan, and DeviantArt were temporarily offline, and some government agencies in the Netherlands and Finland also shut down sites as a precaution; Cloudflare CEO Matthew Prince said attack attempts rose to hundreds of thousands within two hours of disclosure. Security firm Wiz said the vulnerabilities had existed for about 20 years, and OpenAI said researchers used GPT-5.6 Sol to find the vulnerability before attackers and provided detection tools.
#Safety / alignment#OpenAI#English text is being prepared.
METR and Redwood Research said on Thursday that from July 21 to August 11 they conducted a pilot review of OpenAI's internal monitoring systems, with one employee from each organization granted access to OpenAI's internal codebase, documentation, and some technical staff. The evaluators found several new vulnerabilities, including an exploit that allows an agent to evade monitoring of hundreds of recent actions; OpenAI said it has fixed the relevant vulnerabilities affecting production safeguards. The two parties released only a high-level public statement and submitted a 61-page internal report to OpenAI.
#Safety / alignment#Agents#OpenAI
Meta's stock fell about 6% on Thursday, with third-quarter free cash flow at negative $2.12 billion, and it raised the lower bound of its full-year capital expenditure range from $130 billion to $135 billion, making the range $135 billion to $145 billion. The company's third-quarter revenue was $64.6 billion, up 26% year-over-year, above analysts' expectations of $63.37 billion; it also announced the establishment of non-voting Class C common stock to maintain its founder-led structure and finance growth. Its five-year credit default swap spread hit a record high that day.
#Industry news#English text is being prepared.#Meta
OpenAI and Google Cloud have officially announced their partnership: GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna, and Codex will launch on Google Cloud. The first batch of models is already available in private preview to select enterprise customers, with general availability plans to roll out in the coming weeks. Enterprises can call these models via the Google Cloud API, or integrate Codex CLI and desktop workflows into the Gemini platform, with costs consolidated into a single Google Cloud bill. Following Microsoft and OpenAI's April 27 agreement amendment removing Azure's exclusive distribution restriction, and GPT and Codex's arrival on Amazon Bedrock on April 28, OpenAI has now covered all three major U.S. cloud providers within six months.
#Industry news#OpenAI#Google
OpenAI and Google-owned Wiz disclosed that researchers used AI to discover a severe security vulnerability in the PHP core engine, involving memory leaks and unauthenticated remote code execution, with a maximum severity score of 9.2 and affected versions dating back to PHP 5.4. The NSA has issued an emergency advisory requiring relevant organizations to update as soon as possible, and advising that systems be temporarily taken offline if they cannot be updated immediately. The PHP Foundation has released PHP 8.5.12, 8.4.27, 8.3.36, and 8.2.35 to fix the vulnerability. The Wiz research used the Wiz Atlas security agent, Gemini 4 Argon, and GPT-5.6 Sol.
#Safety / alignment#OpenAI#Google
Tesla's total revenue for the third quarter of 2026 was $28.046 billion, essentially flat compared with the same period last year, while net profit was $787 million, down 43% year over year. Musk said on the call that the company is making every effort to push for FSD approval in more markets and hopes to obtain EU-wide approval in the first quarter of 2027, while CFO Taneja said the company is working hard to secure FSD approval in China. Tesla has begun early production of Optimus at its Fremont factory in California; Musk said it may take a year or slightly longer to reach a production rate of 20,000 units per week, and the company plans to integrate the AI5 chip into the robot before April next year.
#Industry news#English text is being prepared.#Embodied AI
At the 2026 World Physical AI Conference and Unitree Partner Ecosystem Conference, Unitree Technology founder Wang Xingxing said that since its release in May 2024, the Unitree G1 robot has cumulatively rolled off the production line about 10,000 to 20,000 units, making it the world's best-selling humanoid robot. The G1 is priced starting at 99,000 yuan and comes in a basic version and a G1 EDU version; it weighs about 35 kg, stands about 127 cm tall, has 23 to 43 joint motors, runs on the UnifoLM large model, and has a battery life of about 2 hours. On October 21, Unitree also launched the H2-X "Kung Fu Legend" humanoid robot, releasing a video showing various martial arts moves and application scenarios such as rescue.
#Industry news#Embodied AI
US President Trump posted on Truth Social, calling on data center builders and major AI cloud providers to pay revenue dividends to their host communities to address rising energy costs and make data centers more readily accepted by local communities. He also demanded that data centers, in addition to generating their own power, must supply surplus electricity to local communities free of charge or sell it to them. Reports say communities across the US continue to protest AI data center projects over rising electricity prices, noise, water shortages and other issues. Hazel Township in Pennsylvania once proposed paying $10,000 in compensation to residents who agreed to have data centers built, but many residents still opposed it.
#Industry news#English text is being prepared.#OpenAI
The UK National Cyber Security Centre issued an alert stating that the PHP scripting language, widely used in website development, has two severe security vulnerabilities; attackers can trigger memory exhaustion to cause memory corruption and, under specific conditions, remotely execute code. The PHP development team released an emergency security update on the 21st to fix the vulnerabilities. The UK Office for Budget Responsibility, Finland's government public network services, and the Netherlands' cybersecurity agency, among others, temporarily took their websites offline out of security concerns. Cloud service vendors such as Tencent Cloud and Alibaba Cloud have issued vulnerability notices, reminding users to investigate and update as soon as possible or take mitigation measures; BBC reported that these vulnerabilities were discovered by researchers at cybersecurity firms with the help of artificial intelligence tools, and had existed for more than a decade without being detected.
#Safety / alignment#Open-source ecosystem#English text is being prepared.
The European Publishers Council (EPC) and organizations including Germany's BDZV and MVFP have asked the European Commission to immediately review Meta's new commercial terms, which take effect on October 30, and to prohibit their application until the proceedings conclude. These organizations argue that the new terms allow Meta to automatically crawl and index the content of websites linked to by users sharing links on the platform, for use in commercial products, which constitutes a violation of the Digital Markets Act. Meta last week removed wording from the terms that explicitly mentioned AI model training, but publishers point out that automated uses such as semantic search, recommendation systems and targeted advertising remain covered; the UK's CMA also plans to summon Meta over the matter.
#English text is being prepared.#English text is being prepared.#Industry news
Fortinet announced unscheduled security updates for FortiOS, FortiManager, FortiAnalyzer, FortiMail, and FortiAuthenticator to fix critical and high-risk vulnerabilities discovered by the Frontier-AI model during security testing, and urged customers to install them as soon as possible. Affected versions include FortiOS 8.0.2, 7.6.8, 7.4.13, 7.2.14, and others; FortiSASE will update automatically by this Thursday. Fortinet also warned that similar AI capabilities could be used by attackers to reverse-engineer patches and construct working exploits within hours or days.
#Safety / alignment#Industry news#Anthropic
The PHP project fixed two high-severity vulnerabilities, CVE-2026-109618 and CVE-2026-109619, in versions 8.2.35, 8.3.36, 8.4.27 and 8.5.12. Security firm Calif found a remote code execution path using GPT-5.6 Sol in OpenAI's Patch the Planet program and reported it on August 11; PHP maintainers said signs of the memory exhaustion issue had already been found during a 2024 audit, but its full impact was not clarified until August this year. Chainguard CEO Dan Lorenc published container images containing the fix more than six hours before the coordinated release on October 21; the PHP Foundation said he unilaterally broke the embargo, while Lorenc said that at the time some vendors had already reported signs of exploitation in the wild.
#Safety / alignment#Open-source ecosystem#OpenAI
On October 21, 2026, two high-severity PHP vulnerabilities were publicly disclosed and patches were released, affecting PHP versions 5.4 through 8.5.11 and potentially enabling remote code execution and full system compromise. Calif researcher Ryan Austin discovered the vulnerabilities within days using AI models, and The New York Times and The Wall Street Journal reported that the initial discovery used OpenAI's GPT-5.6 Sol. Cloudflare CEO Matthew Prince said that after the vulnerabilities were disclosed, attack attempts rose from several thousand to several hundred thousand within two hours, and have now reached millions, with cybersecurity agencies in multiple countries having issued advisories.
#Safety / alignment#Open-source ecosystem#English text is being prepared.
#Industry news#English text is being prepared.#English text is being prepared.
In its quarterly filing, Meta disclosed total future spending commitments of about $825 billion, including $355.7 billion in leases for data centers and energy facilities and $454.5 billion in non-cancelable purchase commitments, nearly doubling within half a year from $182.9 billion and $237.7 billion at the end of March this year. In the third quarter, Meta's capital expenditures of $34.91 billion exceeded operating cash flow of $32.79 billion, and free cash flow for the quarter was negative $2.12 billion, compared with positive $10.63 billion in the same period last year. The company also raised its full-year capital expenditure forecast and announced the creation of non-voting Class C common stock that can be used for capital raising.
#Industry news#English text is being prepared.#OpenAI
Laurent Gisel, head of the ICRC's legal warfare team, told 404 Media during a UN meeting that AI is already selecting targets on the battlefield and causing deaths, and is amplifying destruction, lawbreaking and humanitarian harm. He noted that AI accelerates target generation and the pace of strikes, making it hard for humans to maintain effective control, and argued for prohibiting AI-enabled systems that remove humans from the kill chain as well as autonomous weapons that directly target people. Gisel also mentioned that a draft treaty on autonomous weapons was weakened by Russia and the United States in September this year, and opposed allowing AI into nuclear command and control.
#Safety / alignment#Industry news#English text is being prepared.
#Safety / alignment#English text is being prepared.#OpenAI