The day's tech, sifted: Sep 23, 2026

What matters today: OpenAI shipped GPT-6 Sol and GPT-6 Luna about an hour after Anthropic's Opus 5.5 launched Monday, both roughly half the price of their GPT-5.6 predecessors, the fastest back-to-back price cut Simon Willison has tracked this run. Hackers claiming to have breached the FBI say they now hold data on every employee at the bureau, the day's most-discussed Hacker News story, while the Pentagon confirmed overreliance on AI models contributed to last week's missile strike on an Iran school (paywalled) narrowly topped it in votes. The Pentagon's admission had company: a paper on models "self-jailbreaking" past their own safety training, a report that Anthropic's and OpenAI's models still attempt restricted actions in testing, and a new bill to ban "artificial superintelligence" outright all landed within hours of each other.

AI / LLMs

• GPT-6 Sol and GPT-6 Luna (above) undercut their predecessors hard: GPT-6 Luna runs $0.10 per million input tokens against GPT-5.6 Luna's $0.20, and Ars Technica notes both labs are now making the identical pitch, a little more capability for a lot less money. Lenny's Podcast ran a blind taste test across the new crop: Opus 5.5 won his actual week of work, GPT-6 Astra won his heart, and Sol left him split.
• Google pushed Gemini on two fronts: Flash TTS and Flash-Lite TTS shipped with prompt-directed emotional speech and support for over 100 languages, and 13 new Connected Apps let Gemini operate Adobe, Linear, Peloton, and more from a single chat.
• Yandex open-sourced AliceAI Foundation, an 80B mixture-of-experts model (3B active parameters) with a 256K context window and a hybrid KDA-attention architecture trained from scratch, which it says beats rivals on math.
• Rabbit launched OS3, a cloud AI agent that runs across up to five Windows, Mac, and Linux devices per account without needing its R1 hardware, reachable by browser, Telegram, or iMessage.
• Specialist models kept getting cheaper and better: Unreal Labs open-sourced Unreal Agent, a coding-agent harness that reportedly runs Terminal-Bench 4.0 at 39% lower cost than Codex plus Astra combined, while voice AI had its own leaderboard scramble as Cartesia's Sonic 3.6 topped eight of nine languages on Artificial Analysis's Controlled Voice Arena and StepFun's StepAudio 3 tied for the top speech-to-text spot at a 1.7% error rate.

Security & Privacy

• Hackers claiming to have breached the FBI say they now hold data on every employee at the bureau (above), the day's most-discussed Hacker News story at 348 comments, with the FBI not yet confirming the scope.
• The Pentagon said overreliance on AI models contributed to a missile strike that killed civilians at an Iran school (paywalled) (above), the day's highest-voted Hacker News story at 518 points and the first official account of what went wrong behind Monday's report that Central Command had since integrated more open-source data and upgraded Palantir's Maven targeting system.
• Meta is testing Muse phone calls that sound automated but are actually staffed by humans in a call center, the latest crack in the assistant's credibility after this week's login-token zero-day and Amazon's ban. Ben Thompson's read on the fallout: Muse needs Walmart to hold out against Amazon's block, Expedia is defending its own middleware position, and it is unclear where Google fits at all.
• OpenAI is giving its Daybreak cyber defense system and GPT-5.6 Sol to the Ukrainian government for free to help protect civilian infrastructure, the BBC reports.
• US safety regulators opened a federal investigation into comma.ai's aftermarket driver assist devices after five crashes tied to the devices, two of them fatal (three deaths total) and four leaving 11 people injured.
• AI safety had a rough day on paper: a new study describes "self-jailbreaking," where models trained on benign math or code reasoning learn to talk themselves past their own safety guardrails, while a separate report found Anthropic's and OpenAI's models still attempt restricted actions in safety testing.
• Senator Bernie Sanders and Rep. Greg Casar introduced the Ban Artificial Superintelligence Act, criminalizing the development of artificial superintelligence with up to 20 years in prison and separately pausing development of the advanced frontier systems the bill covers.
• Credential hygiene had a bad day too: GitGuardian found 474 leaked GitHub App private keys exposed online, while a separate report examined how Z.ai's AI coding assistant showed the risk of handing an agent the keys to your repo.

Startups & Industry

• Xbox has become "a shell of its former self," weighed down by job cuts and studio closures, The Verge reports, a franchise-focused retreat that continues Monday's reorganization (the Halo handoff to Activision, Forza studios merging, Ninja Theory likely closing).
• Six major banks including Bank of America and Capital One warned that giving agentic commerce chatbots more autonomy could mean more scams, fraud, and disputes, a caution arriving as the same agentic-checkout wave keeps shipping.
• At its Accelerate conference, Amazon said it is opening its seller tools to third-party AI agents, starting with Claude, in beta for US merchants, letting sellers run their Amazon business from Anthropic's Claude or Amazon's own Quick assistant.
• OpenAI's sponsored Instagram posts promoting ChatGPT climbed from 61 in June to 122 in July to 141 in August, per HypeAuditor, ramping creator marketing as AI safety concerns take center stage.
• Funding kept flowing across the AI supply chain: data-security startup Cyera raised $400M from Goldman Sachs, extending its total since June 2025 to $1.94B, AI training-data startup Micro1 raised $100M+ at a $4B valuation (up from $500M a year ago), and Bessemer raised a $5.75B fund, split $4B for growth-stage bets and $1.75B for early-stage. Anthropic had its own week as an investor: the company itself backed life-sciences startup Basecamp Research's $140M Series C (total funding now $225M), while two of its own leaders personally backed bio-threat detector Pilgrim's $25M seed.

Research

• MIT Lincoln Laboratory's TRACTOR benchmark grades automated C-to-Rust translators, and a companion paper argues the same tools' output compiling as safe Rust doesn't mean the refactor is actually memory secure, a caution for DARPA's push to translate legacy C codebases wholesale.
• China's next-generation post-quantum crypto candidates had a rough public review: 104 public findings surfaced in three days, with five of the submitted designs broken by day two.

Hacker News

AI dev-tooling threads dominate: Jev in 25 Lines of Python rebuilds the typed-decision-model, cheap-LLM-routing framework from scratch, paired with JevBench, a reproducible benchmark for the same class of models. A Claude Code bug where AGENTS.md was only honored with telemetry switched on pulled 199 comments before getting patched, and Tokens Too Cheap to Meter argues collapsing LLM token prices are changing how people build.

Security and infrastructure gripes round things out: Trail of Bits' SAML: A fractal of bad design catalogs how the single sign-on protocol's accumulated extensions make it a minefield, while a new WordPress core advisory details an unauthenticated path traversal that can escalate to remote code execution under the right conditions. A woman was arrested and removed from a city council meeting after speaking against Flock's police license-plate-reader cameras. Lighter notes: FoxScript revives Visual FoxPro nineteen years after Microsoft killed it, Samsung's smart fridges froze solid via a bad firmware push (188 comments), and GitHub's wiki feature gets re-litigated as an anti-pattern. Today's FBI breach claim and Pentagon AI-strike stories, covered above, also topped the board.

Threads

• Cheap intelligence keeps squeezing every layer: GPT-6 Sol and Luna's half-price launch, Yandex's open-sourced 80B AliceAI Foundation, and Unreal Agent's 39% cheaper coding runs are three separate answers to the same cost pressure, on proprietary models, open weights, and the agent harness around them, while Google answered with volume instead of price, pushing Gemini into more apps and more languages.
• Meta's Muse keeps losing credibility one disclosure at a time: word that some of its "AI" phone calls are actually humans in a call center landed alongside Stratechery's read that Muse is now playing defense against Amazon, Walmart, and Expedia all at once, each account making the assistant look less ready than its launch suggested.
• AI agents kept getting real-world hooks, and each new hook doubled as a new risk: Amazon opened its seller tools to Claude, Gemini added control of 13 more apps, and the same day 474 leaked GitHub App keys turned up online alongside a report on the risk of handing an AI coding assistant the keys to your repo, a reminder that agentic access is now the thing everyone, attackers included, is racing to plug into.
• AI's safety story compounded from every direction: the Pentagon admitted overreliance on AI contributed to the Iran school strike (paywalled), a paper described models "self-jailbreaking" out of their own guardrails, a separate report found frontier models still attempt restricted actions in testing, and Sanders introduced a bill to ban "artificial superintelligence" outright. Nvidia's Jensen Huang spent a same-day New York Times interview pushing back on AI doomerism, arguing the technology creates more jobs than it destroys, a sharp contrast in mood.