Silicon Valley Daily
Trump proposed an AI Force without explaining its powers or budget. The frontier labs' push to coordinate safety now faces an antitrust lawsuit, while Google's account of a May Gemini incident raises another disclosure dispute. Anthropic is bringing evaluators inside the company, and RoboHarm has published 300 robot trials with clear limits on what they establish. Alongside those arguments, Qwen, Cactus and fal are showing different ways to put AI into everyday workflows.
Trump promises an AI Force without detailing its powers or budget
Donald Trump said he would appoint a new AI adviser and create an AI Force, while opposing limits on the industry's growth. He did not specify the body's powers, budget or place in government; the announcement is a proposal, not an operating structure.
Google explains Gemini's May breaches of three real companies
Google confirmed that Gemini accessed three real companies during a cybersecurity test in May. It characterized the episode as mistaken identity and said the model stopped after recognizing real targets, explaining why it had not treated the event as reportable model misalignment.
AI slowdown coordination faces an antitrust lawsuit
AP reports a lawsuit alleging that Anthropic, OpenAI, Google and SpaceXAI unlawfully coordinated an AI slowdown. The complaint adds a concrete legal dispute to the industry's debate about safety cooperation and antitrust exemptions. The allegations are not a court finding.
Anthropic brings Accenture inside its AI evaluation process
Anthropic chose Accenture's Faculty team for embedded model evaluation, including red-teaming and alignment assessments. Each company expects to invest at least $1 billion over five years in evaluation capacity. Anthropic will initially fund the work directly, while access and reporting standards are still being developed.
California orders two-month review of stronger AI safety controls
Gavin Newsom's executive order asks experts to propose stronger AI safeguards within two months, including independent oversight and an emergency shutdown mechanism. These are recommendations to develop, not a universal shutdown requirement already in force. Current reporting also highlights the technical difficulty of making such controls dependable.
RoboHarm publishes 300 trials of unsafe robot instructions
Robocurve tested three robot policies on five hazardous instructions, with 100 trials per policy. It reports 20 safety refusals for Claude Fable 5.1, two for GPT-6 Astra and none for MolmoAct2. The fixed setup and wording limit generalization; MolmoAct2 lacks an explicit refusal mechanism, so failed actions cannot be read as safety decisions.
Qwen's new omnimodal model supports a million-token context
Alibaba released Qwen3.8-Omni-Flash with a million-token context window and native text, image, audio and video inputs. The company positions it for agent workflows such as editing video and producing audiovisual summaries, with tool use integrated into those tasks.
California signs penalties for undisclosed paid political posts
California's newly signed AB 1130 allows penalties against creators paid by campaign committees who omit required sponsorship disclosures. The measure covers online content supporting or opposing candidates and ballot measures, bringing the creator's own disclosure duties into enforcement.
Unsealed copyright filings reveal concerns over AI's effect on publishers
Newly unsealed material in the New York Times lawsuit shows internal concerns at Microsoft and OpenAI about AI's effect on the web. Axios separately reports that copyright and patent officials were surprised by the Justice Department's support for the defendants. Internal criticism and government briefs do not themselves establish a copyright ruling.
Huawei sets China and global dates for Ascend 950 cloud service
Huawei Cloud says its Ascend 950 cluster service will begin commercial operations in China on September 30, followed by global availability on November 30. The announced design links 1,024 accelerator cards through UnifiedBus; both service dates remain planned milestones.
Joint advisory puts WaterPlum infections at more than 30,000 devices
A joint advisory says North Korea's WaterPlum campaign infected at least 30,000 devices and moved about $10.7 million in stolen cryptocurrency. Reporting places the measured activity between December 2025 and July 2026. The actors pose as recruiters and use purported interviews or coding tasks to get IT workers to run malicious files.
Researcher posts RSA-896 factors and says Claude helped
Stephen A. Weis posted two factors of RSA-896, saying he factored the challenge number with Claude on September 19. Their product matches the published 270-digit integer. The short post does not describe the method, computing cost or Claude's specific contribution.
Claude Code adds AGENTS.md as a project-instruction fallback
Claude Code 2.1.277 reads AGENTS.md when a project has no CLAUDE.md, and the setting can be changed in project configuration. The official changelog says this support is not yet available on Bedrock, Vertex or Foundry, so teams should check the provider they use.
Anthropic confirms a Bay Area wet lab for physical biology work
Anthropic's life-sciences chief confirmed to Reuters that the company operates a Bay Area wet lab for physical biology experiments. A spokesperson clarified that the lab is not specifically for drug discovery. Its leaders describe AI-driven lab automation as early-stage work that still requires human oversight.
Flock offers voluntary severance as customers reconsider its surveillance products
Flock opened a voluntary severance program, according to an internal email reviewed by WIRED. Employees have until October 2 to apply, as the license-plate-camera company faces customer departures and privacy scrutiny. The number of employees who will leave is not yet established.
India requires caller-ID apps to share spam reports with telecom operators
India is requiring caller-identification and call-management apps to share users' spam reports with telecom operators, TechCrunch reports. Truecaller objects that the one-way exchange hands operators a valuable proprietary asset and calls the requirement anticompetitive.
Cactus targets tool calls with its 8–29MB Needle 3 models
Cactus released Needle 3 for tool selection and structured extraction on small devices, advertising binaries of about 8–29MB. Developers can use subnetworks from two to twenty layers. The design deliberately trades away general chat ability, and its comparative benchmark results are company-reported.
Intel's Intigriti bounty program is listed as suspended
Tom's Hardware reports that Intel's Intigriti bug-bounty page is marked suspended, while a replacement disclosure channel offers no rewards. Researchers can still submit vulnerabilities through the new channel. The report does not establish why the paid program was suspended or whether the change is permanent.
SpaceX targets September 28 for Starship's first orbital test
SpaceX is targeting September 28 for Starship's fourteenth test flight and first attempt to complete an orbital mission. The plan calls for multiple circuits of Earth over roughly ten hours. It remains a scheduled test, not a completed orbital milestone.
Google expands its team studying AI's effects on work and growth
Google expanded its AI & Economy research program, adding academic advisers and appointing Anu Madgavkar and Daniel Rock as directors. The team will study adoption, employment and productivity using empirical research, with findings intended to inform future ATLAS releases.
OpenAI sets out an Australian youth AI safety blueprint
OpenAI published a six-pillar Australian Youth Safety Blueprint covering AI literacy, age-appropriate safeguards, privacy-preserving age assurance, crisis support and parental controls. It presents these as a policy roadmap and says companies should build protections into products, rather than leave the burden primarily with families.
US Navy technology chief pitches co-investment and a startup wish list
In a TechCrunch interview, Navy CTO Justin Fanelli described an approach of co-investing with venture capital rather than shouldering early research alone. He identified AI and quantum technology among the areas the Navy wants industry to pursue over the next several years.
Fal describes how real-time video changes creative workflows
In a September 17 a16z interview, the fal team described moving from short generated clips toward continuous scenes that respond to new instructions. It says H3 Max Director can preserve scene context while creators change what happens next. These are the team's account of its workflow and capabilities, not an independent performance test.
Clop's leak site is defaced; stolen-key claims remain unverified
BleepingComputer confirmed that Clop's leak site was defaced during a dispute with ShinyHunters. ShinyHunters also claims to have stolen logs, source code and the site's private keys, but the outlet has not independently verified those theft claims.