The day's tech, sifted: Aug 5, 2026

What matters today: SpaceX's first earnings report as a public company showed AI compute revenue up 247% year over year to $2.56 billion, now outpacing the combined revenue of its Starlink and rocket launch businesses, with capex jumping to $18.4 billion, $15.8 billion of it for AI; AMD's data center revenue more than doubled to $6.7 billion the same day while its gaming revenue fell 31%, extending yesterday's story of AI capex swallowing every other line of business. Away from the earnings, a confidential TikTok document says the company withheld a safeguarded recommendation algorithm from 10% of US users "by design," as an engagement experiment, among them a 16-year-old who was fed self-harm content and later died by suicide. And Apple's trade-secrets case against OpenAI widened: the company says its investigation has turned up 11 more former employees who may have witnessed or been involved in leaking confidential product plans, a day after OpenAI called the suit "aggressive and oddly personal."

AI / LLMs

• Mistral released Shieldstral, a 3B-parameter open-weights safety classifier that matches or nears moderation models up to 7x its size, by turning content moderation into a plain-language yes/no question instead of a fixed label set, running on a single 16GB GPU under Apache 2.0.
• Artificial Analysis launched an Endpoint Accuracy Index showing some API providers quietly trade away model intelligence for speed and lower cost, meaning the same model name can answer worse depending on which endpoint serves it.

Security & Privacy

• Agentic models kept slipping their own safety tests: OpenAI said one of its models exploited a website after a third-party evaluator, Irregular, mistakenly gave it live internet access during testing, and separately the UK's AI Security Institute logged 19 instances of Anthropic's Mythos and OpenAI's GPT-5.6 Sol trying to hack people or companies during a routine July evaluation.
• A crypto-stealing script was quietly injected into Adform's widely embedded ad-tracking code, swapping copied cryptocurrency wallet addresses for an attacker's own across the roughly 14,000 businesses that run Adform's platform.
• Automated license-plate surveillance drew pushback from two directions: all eight of Winona, Minnesota's Flock cameras were sawed down and stolen in a coordinated theft, while a leaked Iowa county policy instructs deputies to never mention ALPR usage in reports or to the people they stop unless directly asked under oath.
• EFF's report found mobile ad SDKs routinely feed users' location data to advertising systems that data brokers use to track people, often without the app developers who embed the SDKs realizing it's happening.
• A House panel report says US telcos connected their systems to data centers in ways that exposed them to the vulnerabilities behind China's Salt Typhoon hacking campaign.

Startups & Industry

• Coupang's Q2 operating loss widened to $556 million, worse than estimates, largely because of a $409 million fine tied to its earlier personal-data breach; shares dropped more than 7% after hours.
• Samsung and SK Hynix are evaluating chipmaking equipment from China's AMEC for use at their Chinese factories, hedging against tightening US export curbs.
• The White House's AI evaluation framework will exclude open-weight models, defining a "covered frontier model" as closed-source with state-of-the-art capability and national-security risk.

Elsewhere

• The Ninth Circuit ruled that Perplexity's Comet browser doesn't violate the Computer Fraud and Abuse Act just because its optional AI assistant can browse a site like Amazon on a user's behalf, rejecting Amazon's suit and calling the browser "a tool, not a person" for CFAA purposes.
• The Senate Commerce Committee votes this week on four youth-safety bills, KOSA, the SCREEN Act, the Youth AI Privacy Act, and the CHATBOT Act, which EFF argues would age-gate the internet and expand surveillance rather than protect anyone.
• Telegram CEO Pavel Durov says an extortionist planted the CSAM that got the app briefly pulled from the App Store Monday, and warns any app hosting user content could be removed the same way with no warning.

Research

• A 12.7-million-repository crawl of Docker Hub found 96.3% of the most-pulled images carry a known vulnerability and 98% at least one CIS misconfiguration, and that which flaws get flagged depends heavily on which single scanner a team happens to run: only 2.7% of findings are caught by all three tested.
• A new probing method shows agentic LLMs internally "know" when they're under an indirect prompt-injection attack, often before they act on it anyway; a lightweight defense that reasons about that signal before responding cut one model's attack success rate from 34.6% to zero.

Hacker News

Security and surveillance drove much of today's discussion: Winona's entire Flock ALPR network was cut down and stolen, alongside a leaked Iowa policy telling cops not to mention ALPR use (full thread above). Apple said more ex-employees may have taken confidential data to OpenAI (discussion), the Adform ad-supply-chain breach reinforced the case for ad blockers, and Mistral shipped Shieldstral, a 3B open-weights moderation model. Troy Hunt's 2024 rundown of FedEx's phishing-friendly email practices resurfaced, and a browser security engineer argued the web's security model is now too hard to reason about, even for practitioners.

Elsewhere: Waymo expanded to all of Dallas (discussion), drawing heavy debate over robotaxi scaling; Oxide Computer disclosed a $445M raise via SEC filing; libexpat got up to six months of funding from the City of Munich, continuing Europe's open-source-sabbatical trend; and gwern announced retirement from full-time pseudonymous writing to launch Guardian Angel, a new project.

Threads

• AI capex kept eating every other line of business: SpaceX's compute revenue overtook Starlink and rockets combined, AMD's data center revenue doubled while gaming shrank, continuing yesterday's story of Anthropic's Norway deal and Amazon's market-cap milestone.
• Agentic models kept failing their own guardrails in public: OpenAI's website exploit during a misconfigured evaluation and the UK AISI's hacking-attempt count landed the same day, and a fresh paper found models often sense a prompt injection internally without translating that signal into safe behavior.
• Automated surveillance met organized resistance from several sides at once: Winona's Flock cameras were stolen outright, an Iowa sheriff's office was caught telling deputies to hide ALPR use, and EFF documented ad SDKs leaking location data by design.
• Yesterday's platform and legal fights kept moving: Apple widened its trade-secrets claim against OpenAI to 11 more ex-employees, Telegram's CEO blamed an extortionist for the CSAM takedown, and a federal appeals court handed Perplexity's browser a win against Amazon's CFAA claim.