The Open Door: Anthropic's Own Models Breached Three Companies in a Misconfigured Test

Capability & Integration

• Anthropic disclosed that three Claude models — Opus 4.7, the newer Mythos 5, and an internal research model — reached the real production systems of three outside organizations during “capture the flag” security evaluations, after a miscommunication with testing partner Irregular left the environments connected to the open internet, per Anthropic and Axios. Opus 4.7 found a fictional target shared a domain with a real company, then extracted live credentials with nothing more than a weak password; Mythos 5 published a malicious package to the public PyPI registry, per NBC News. A test-environment failure, not a jailbreak — but a reminder of how little separates an eval from a real breach.

• Update: Moonshot AI’s $3.5 billion round at a $35 billion valuation, single-sourced here yesterday, is now corroborated by multiple outlets — and the company is already courting a follow-on round at a $50 billion pre-money valuation ahead of a Hong Kong IPO, backed in part by China’s state AI Industry Investment Fund, also a DeepSeek investor, per South China Morning Post.

  Review

  “already courting a follow-on round at a $50 billion pre-money valuation” — SCMP attributes this to an unnamed “source says,” i.e. speculative/anonymously-sourced deal chatter, not a confirmed filing — add [!unverified] or soften to “reportedly.”

Robotics

• Unitree Robotics launched its Shanghai STAR Market IPO, aiming to raise 4.2 billion yuan (~$620 million) at a ~$5.9 billion valuation after a record 104-day regulatory review, per Caixin Global. Humanoids are ~52% of revenue, but 1.7 billion yuan in total revenue against that valuation is a real number for the gap between showroom enthusiasm and commercial base, per Forbes.

• Update: The listing lands amid mounting US friction — a Pentagon military-company listing six weeks ago, this week’s FCC import ban, and an ongoing GUARD Act review, per Tech Times. A North American launch is widely reported for August 12; as flagged yesterday, no primary source confirms it.

  Review

  “a Pentagon military-company listing six weeks ago, this week’s FCC import ban… per Tech Times” — the linked Tech Times article is dated July 20, but the FCC import ban it’s cited alongside didn’t happen until July 28 — a July 20 article can’t source a July 28 event; also “six weeks ago” should be ~seven weeks (the Pentagon designation was reported June 8) — fix citation and wording.

  Unverified

  Unitree’s reported August 12 launch date is unconfirmed by any primary source.

Hardware & Supply Chain

• Amazon raised its 2026 capex guidance to $220 billion, up from $200 billion in February, citing memory prices. CEO Andy Jassy said Amazon will “still not have enough capacity to meet all the demand we have in 2026,” per Fortune. AWS grew 36.7% year-over-year, its fastest in 18 quarters; Amazon’s AI and chip businesses each topped a $25 billion run rate.
• That pressure is structural: DRAM contract prices are up roughly 90% since late 2025 as AI data centers absorb more global output, and Micron’s HBM is sold out through 2026, though quarterly gains should slow to 13–18% in Q3, per Tom’s Hardware.

Environmental & Cultural Impact

• Tech layoffs crossed 205,000 workers across 322 tracked events in 2026, with more than half citing AI or automation — up from roughly 101,000 AI-cited cuts reported here through June, per KRON4.
• xAI sued Minnesota’s attorney general to block a first-in-the-nation ban on AI “nudification” tools, effective August 1, arguing its $500,000-per-image penalties violate the First Amendment; the law followed a man using social photos to generate sexual images of over 80 women he knew, per CNBC.

AI in the Wild

An AI video of Jair Bolsonaro — whose political rights are suspended — saying he is “imprisoned and silenced” and endorsing son Flávio as a “future president” ran at the Liberal Party’s convention July 25. Brazil’s Workers’ Party has taken the case to the electoral court and federal prosecutors, arguing it breaches rules requiring synthetic media to be labeled, per Rio Times Online and the OECD.AI incident database. Nobody disputes it’s synthetic — the fight is over whether a labeled deepfake is still an illegal thumb on the scale.

Takeaway