The day's tech, sifted: Aug 31, 2026

What matters today: Over 12TB of Steam's old "Steam2" content servers leaked, covering every build uploaded between 2003 and 2013, including abandoned Half-Life 2: Episode 3 assets and cut Portal 2 content. Separately, Anthropic is signing out some Claude users, clearing their saved payment methods, and issuing refunds after infostealer malware on their PCs hijacked login sessions to drain paid usage. And in Hong Kong, Shein priced its IPO at $1.7B, a $26B market cap far below its ~$100B valuation in 2022, as tariffs and competition squeeze growth.

AI / LLMs

• An NPR analysis found popular chatbots debunked false narratives pushed by Russia, China, and Iran most of the time; Google's AI-generated search summaries did too, but at a lower rate than dedicated chatbots.

Devtools & Infra

• MiniMax's H3 and H3 Max video models are 50% off on Vercel's AI Gateway through September 13; the discounted model IDs are unchanged, so existing integrations pick up the rate automatically.

Security & Privacy

• Anthropic is signing out some Claude users, clearing saved payment methods, and refunding drained usage after infostealer malware on affected PCs stole active login sessions and used them to run up charges.

Startups & Industry

• John Ternus formally becomes Apple's CEO on September 1; Bloomberg's rundown of what he inherits adds rising component costs and staff retention to yesterday's expected executive reshuffle, with Tim Cook staying on to handle Trump and China in the transition.
• Shein's Hong Kong IPO raised $1.7B, valuing the fast-fashion retailer at $26B, a fraction of the ~$100B it commanded in 2022, as tariffs and slowing growth weigh on the business.
• ING estimates the US-led AI boom accounts for roughly a third of recent US economic growth, offsetting a global growth squeeze from the energy crunch even as its outsize role worries economists.
• Taiwanese prosecutors raided PCB maker Unimicron, a key supplier to Nvidia and Intel, alleging it mislabeled China-made parts as Taiwan-made.
• Chip design startup Agentrys raised a $19.1M seed led by Etna Labs plus a $5.4M pre-seed led by MediaTek; the Nvidia-alum-led startup lets users build their own chip-design agents.

Research

• ROPE proposes routing agent tool calls through origin-tagged policy checks to block indirect prompt injection, addressing cases where a malicious instruction only reveals itself once tool parameters are set at runtime, past where static screening can catch it.
• ContextLeak formalizes how a malicious tool can exfiltrate an LLM agent's runtime context: the user's prompt, its execution trajectory, even its tool list, by getting the agent to pass that context as tool input, the step prior work had mostly skipped over.
• A mechanistic-interpretability study traces the internal circuits a safety-aligned LLM uses when it gets jailbroken, isolating the subnetworks in Llama-2-7B-chat that generate compliant responses to adversarial prompts.

Elsewhere

• NASA administrator Jared Isaacman briefly stepped away from Sunday's Roman Space Telescope post-launch briefing to take a call from President Trump, who then addressed the room to congratulate the mission team.
• Google Maps now shows "Lake America" for US users and "Lake Ontario" for Canadian users, and both names for everyone else, citing the US government's own name change.

Hacker News

Two threads worth grouping: institutional and coordination pieces, and a lighter historical and language pair. Startup Anti-Patterns (110 points, 58 comments) and Coordination Headwind (138 points, 45 comments) both dig into why organizations struggle to move coherently, one from a founder's-eye view, one via a biology metaphor. Bigger and more contentious: algorithmic rent-pricing litigation (105 points, 77 comments) tracks how state and local laws are catching up to pricing algorithms, and a climate piece arguing the timeline is tighter than assumed (paywalled, 112 points, 138 comments) pulled the most replies of the non-top story.

Top of the pile by a wide margin: a piece on deliberate word choice, "I just chose words carefully", at 438 points and 109 comments, well clear of everything else in the set. Rounding out the day, a Smithsonian piece on Benjamin Franklin's use of pseudonymous alter egos to write and argue more freely (100 points, 73 comments).

Threads

• Two Anthropic stories cut in opposite directions: a Claude account-security incident traced to infostealer malware, and arXiv research (ROPE, ContextLeak) formalizing the same class of agent-context attacks the industry is racing to patch.
• Corporate succession and scale shared the day: Ternus's inherited cost and retention problems at Apple, Shein's valuation collapse since 2022, and Unimicron's supply-chain scrutiny all turn on pressure building under fast-grown companies.
• Old software, new life: a 12TB Steam leak surfacing a decade of shelved game builds landed the same week AI's economic footprint (ING's third-of-growth estimate) keeps compounding on top of everything built since.