Federal investigators probe Comma's hands-off driving tech after fatal crashes

Comma.ai, the startup founded by hacker George Hotz, is facing a federal investigation after five reported crashes involving its aftermarket hands-off driver assistance tech, two of which killed three people. The National Highway Traffic Safety Administration’s Office of Defects Investigation says the crashes involved Comma devices failing to detect or respond to slow or stopped vehicles in the same lane, with as many as 11 people injured across four of the crashes, according to TechCrunch’s account of the investigation.

At least one fatal crash appears to have involved a third-party fork rather than stock software: a February 2026 crash in Louisiana in which a Toyota RAV4 running the FrogPilot fork of Comma’s openpilot struck a stopped, fully marked police unit with its emergency lights on, killing two rear-seat passengers. The other fatal crash could not be immediately identified from the cited reports, and neither Comma.ai nor Hotz responded to requests for comment, so how much of the record reflects modified software remains unresolved.

Featured source: TechCrunch , Comma .

Other Stories

China rejects a US call to slow frontier AI development

Chinese officials, state media, and engineers rejected Anthropic CEO Dario Amodei’s proposal that frontier labs pause their most powerful model work, calling it an attempt to preserve the US lead, with a Foreign Ministry spokesperson on September 11 describing the call as fear-mongering that would undermine global AI governance, according to Rest of World. Amodei’s plan would require cooperation with China alongside chip-sale bans, smuggling crackdowns, and limits on remote datacenter access. The rejection comes ahead of a scheduled September 24 meeting between Presidents Trump and Xi, where the US is expected to raise shared AI risks — and with China having published a third edition of its AI Safety Governance Framework on September 14, the disagreement reads as a fight over terms and control rather than whether AI risks exist.

Filed from: Rest of World , Amodei’s proposal .

Anthropic and OpenAI open a price war on cheaper models

Anthropic released Claude Opus 5.5 and OpenAI released GPT-6 Sol and GPT-6 Luna within about an hour of each other, both pitches centering on lower cost rather than major capability jumps. Anthropic cut input and output prices 20 percent versus Opus 5 and cache reads 60 percent, claiming roughly 40 percent lower cost on typical workloads, while OpenAI halved Sol and Luna API prices versus promotional GPT-5.6 pricing, with Luna at $0.10/$0.50 per million input/output tokens. Independent developer Simon Willison notes the cuts price GPT-6 Sol input the same as Grok 4.7, though he also found Opus 5.5 at maximum reasoning effort can over-think to the point of exhausting its output limit on a toy drawing test.

Filed from: Ars Technica , Simon’s pricing breakdown .

ShinyHunters claims a breach of FBI agent and applicant data

The ShinyHunters extortion group claims on its dark-web leak site to have stolen sensitive data on almost all FBI agents and job applicants, including names, home addresses, and phone numbers. 404 Media, which first reported the claim, says it received a sample and verified a portion against public records, and reports the group says it entered through an Oracle PeopleSoft server before pivoting to an Amazon-hosted government cloud. The FBI did not respond to requests for comment and the claim remains single-sided with partial independent corroboration; the group says the hack is not financially motivated and demands removal of an FBI report it calls false.

Filed from: TechCrunch , 404 Media .

UK police arrest two EvilTokens suspects as Microsoft seizes phishing sites

A Microsoft-led coalition with US and UK law enforcement disrupted the EvilTokens phishing service, seizing 50 operating websites, disabling more than 150 supporting domains, and notifying victims. London’s Metropolitan Police arrested two men, aged 32 and 38, on September 18 as alleged administrators; both were released on bail while the investigation continues. The device-code phishing kit, sold as a service since February, compromised 12,000 inboxes across more than 10,000 organizations and bypassed multi-factor authentication — distinguished by an AI chatbot that analyzed victims’ inboxes to pick targets and fraud strategies. Microsoft says it observed 10 to 15 distinct campaigns launching every 24 hours since mid-March.

Filed from: The Register .

Researchers describe the first Windows implant that lets LLMs pick its next move

Cisco Talos researchers describe CLOSEDQUORUM, a Go-based Windows implant that queries up to four language models — DeepSeek, Qwen, Mistral, and Google Gemini — and lets them vote on which pre-defined post-compromise action to take next, from credential and crypto-wallet theft to injection and persistence. Talos calls it, to its knowledge, the first publicly documented Windows implant to use LLMs this way for command-and-control, removing the human operator as the bottleneck once deployed. Talos says it has not observed the implant in the wild and linked its developer to carding-forum postings dating to 2025, so this is an analyzed sample demonstrating a technique rather than an active campaign; Talos also open-sourced its CAIRN toolkit for hunting AI-integrated malware.

Filed from: The Register , Cisco Talos .

Z.ai apologizes for uploading user code and open-sources ZCode

Chinese AI company Z.ai apologized after developers found its ZCode code-generation tool packaging and uploading entire user workspaces — including complete project histories — encrypted to Alibaba Cloud, with the decryption key held only by Z.ai’s servers. The researcher who flagged it says there was no user option to disable the behavior and no disclosure in the privacy policy. ZCode says the uploaded data was never used to train models, that outside assessors concluded all previously uploaded data has been deleted, and that the Repo Wiki feature behind the uploads was removed; it has open-sourced the project for community scrutiny, with a full security assessment report still pending.

Filed from: The Register .

Ema raises $77M to automate enterprise work with agent teams

Ema, a startup selling teams of AI agents it calls “AI employees” for HR, IT, and finance processes, raised $77 million in Series B funding led by Creaegis, with Accel, Section 32, and Prosus increasing their stakes. The all-primary-equity round brings total funding to $140 million and more than quadruples the 2024 valuation, a figure the company declined to disclose. Founder Surojit Chatterjee says the systems coordinate multiple agents across a company’s existing applications and that customers are on their way to replacing some large SaaS products outright — the founder’s characterization of an ambition that puts Ema against both frontier labs moving into enterprise work and incumbents defending their seats.

Filed from: TechCrunch , Ema .

Privacy group says EU plan lets AI reuse decades of personal data

Max Schrems’s campaign group noyb attacked an EU proposal to adjust GDPR for AI development, arguing a new legitimate-interest basis for processing personal data “in the context” of building and operating AI systems would let large firms reuse decades of collected data with almost no restrictions. Schrems called it “a digital expropriation of Europeans” subordinating privacy rights to AI-company profits. The changes stem from the Commission’s competitiveness program to lighten regulatory load, with the amendment most recently discussed as Article 88c (or 88bis in a leaked Council compromise draft) — wording and adoption still legislative unknowns, with noyb raising the prospect of another court challenge if the text passes as-is.

Filed from: The Register .

Researchers begin a community rebuild of the terminated US climate assessment

After the White House shut down work on the next US National Climate Assessment — cutting research-program funding, removing contributing authors, and taking past reports offline — the American Geophysical Union and American Meteorological Society’s community “US Climate Collection” published its first paper. The paper, by 175 authors across two society journals, is a roadmap prioritizing the research meant to fill the vacuum, from measured trends versus model projections to newer ground such as machine-learning-based climate simulations. The collection will gather peer-reviewed papers as a resource for future report writers rather than producing a single coordinated report, so its value depends on researchers contributing the underlying studies.

Filed from: Ars Technica , AGU/AMS announcement .

Qualcomm launches two flagship Snapdragon chips on 2nm

Qualcomm announced the Snapdragon 8 Elite Gen 6 alongside an 8 Elite Extreme Gen 6, both positioned as flagships on a 2nm process. The company describes the differences as relatively minor — improved AI processing, video capture, and gaming performance on the Extreme — with modems, display support, security features, GPS, and memory support shared. Qualcomm no longer publishes detailed core breakdowns, so head-to-head performance claims await independent testing and shipping phones; the announcement sets the silicon baseline for next year’s Android flagships.

Filed from: The Verge , Qualcomm .

From the Community

BGP Hijack and Forged RPKI Origin Used to Push Malicious Virtualizor Update

Beginning at 20:57 UTC on 28 August 2026, a more-specific prefix covering Softaculous’s update endpoint and client and billing sites was announced with a forged origin of AS24940, and because Hetzner’s ROA for the parent block allowed /24 lengths, the route validated under RPKI and propagated widely. Softaculous says the attacker paired the hijack with a technically valid TLS certificate to deliver a malicious Virtualizor update package to a small number of installations. The APNIC analysis notes the uncontested route let the attacker control a quorum of vantage points, defeating the multi-perspective corroboration Let’s Encrypt uses against localized hijacks; Hetzner has since tightened three ROAs and added an ASPA record, though the same maxLength gap persists across much of its address space. The affected-installation count comes from Softaculous’s own description and has not been independently quantified.

Filed from: APNIC Blog .

Rust Team Discloses GitHub Actions Secret Leak Through Cached Miri Output

The Rust Security Response Team reported that Miri writes all environment variables into target/, so workflows that cache that directory can persist secrets into a cache that pull requests are able to read. Since anyone who has previously landed a change can trigger PR CI, and GitHub sometimes hides overwritten commits, the team says this kind of extraction can be difficult to detect. The short-term fix restricts Miri to preserving CARGO_* variables except CARGO_*_TOKEN, plus OUT_DIR, shipping in the 2026-09-22 nightly; an ecosystem scan found one affected repository and seven that should be cautious. The team describes the underlying behavior as not a vulnerability in itself; exposure depends on a workflow that caches target/ and grants PRs cache read access.

Filed from: Rust Blog .

WordPress Patches Unauthenticated Path Traversal That Can Lead to Remote Code Execution

WordPress 7.1.2 fixes an unauthenticated path traversal in get_page_template() resolution that allows an attacker to include a chosen readable local .php file from outside the active theme directories, which the advisory says can lead to remote code execution when server and theme preconditions are met. Those conditions include a theme with a top-level page-* directory — affecting the legacy Twenty Twelve and Twenty Fourteen themes and third-party themes Neve, Hestia, and Sydney — plus a readable local PHP target file, with the pearcmd.php PEAR-to-RCE path available when register_argc_argv is On. The fix has been backported to all branches back to 4.7.

Filed from: WordPress Security Advisory .

DeepSeek Describes a Sandbox Platform Serving Millions of Agent-Training Environments per Day

A DeepSeek technical report presents DeepSeek Elastic Compute (DSec), a production sandbox platform exposing FnCall, container, microVM, and full-VM backends through a unified SDK for large-scale agentic training and evaluation. DSec coordinates placement and lifecycle across a cluster, composes environments from independently versioned layers, and loads image data on demand from the Fire-Flyer File System (3FS). The report states that one production-scale unit spans roughly 160 nodes and serves about 3 million sandboxes per day, sustaining over 380,000 concurrent sandboxes — figures from a vendor-authored report that have not been independently verified.

Filed from: arXiv .

Samsung Halts Smart Fridge Update in Korea After Units Stop Working

Samsung stopped distributing a SmartThings update to smart refrigerators in Korea after owners reported that units stopped working entirely, spoiling food, and that automatic door opening and interior lighting failed. A Samsung representative said an error occurred during testing of a refrigerator software update and that the company was taking emergency measures through its service centers, while ZDNet Korea attributed the distribution to a system error. User reports describe technicians replacing mainboards, and customers have asked Samsung to compensate them for spoiled food. Details rest on Samsung’s forum statement and ZDNET Korea reporting; no full technical account has been published.

Filed from: Android Authority .