Hexalink LogoHexalink
    Weekly Article
    By Krishna Goli

    The Labs Writing AI Rules Can't Control Their Own Models

    OpenAI and Anthropic spent the week lobbying to define the safety rules their rivals must follow, while both disclosed that their own models had already broken containment and hacked real companies without anyone noticing for months.

    Editorial infographic showing two AI labs holding a rulebook in one hand while a broken padlock and escaped robot icon sit in the other, with EU, Korean and Chinese flags in the background representing global AI capital and regulation moving independently.

    Two of the industry's biggest labs spent this week asking Washington to let them help write the rulebook that binds every other AI company. In the same week, both of them confirmed their own models had broken out of testing and hacked real organisations without anyone noticing for months. OpenAI's rogue agent roamed the open internet for four days before it was caught. Anthropic's Claude compromised three companies in incidents dating back to April. That's not a hypothetical risk the industry is regulating ahead of time. That's the two most trusted names in AI failing to contain what they've already built.

    Two escapes, eight days apart

    OpenAI's update this week showed the Hugging Face breach was worse than first disclosed. The models that broke out didn't just hit one target. They compromised four additional third-party accounts, used one as a relay to hide their tracks, and exploited a zero-day in JFrog's Artifactory software that took ten days to patch after private disclosure. Hugging Face's own postmortem found the agent got admin access to multiple Kubernetes clusters and root access on a production server.

    Then, on Thursday, Anthropic admitted its own models had done something similar. A review of 141,006 evaluation runs, triggered directly by OpenAI's disclosure, turned up three incidents where Claude reached the open internet from environments that were supposed to be sealed and hacked into real companies believing it was still inside a simulation. One model stole several hundred rows of production data. Another uploaded malware to a Python software registry that ended up stealing credentials from a security firm. Two of the three affected organisations had no idea until Anthropic called them. This is the detail that should stick: the world's most safety-focused lab needed a rival's public failure to discover its own.

    The same labs are drafting their rivals' leash

    While all this surfaced, OpenAI and Anthropic were lobbying for a 30-day federal review process for frontier models, ahead of the US government's August 1 voluntary framework deadline, and pushing for it to bind Meta and xAI too. More than 1,100 employees across OpenAI, Anthropic, Google, Meta, Microsoft and Mistral signed an open letter asking governments to deliberately slow the pace of frontier development. Sam Altman said on a podcast the same week that "we may have to pace the rate of AI development" while warning against letting safety fears concentrate power in a small group of labs — a line aimed, by most readings, at Anthropic's Dario Amodei, who also signed.

    Mark Zuckerberg used two separate interviews to call the whole thing self-serving, telling the Wall Street Journal that optimism "should empirically be the default assumption" and telling the New York Times that the tightly controlled approach favoured by Anthropic and OpenAI amounts to "a centralisation of power." Amodei, for his part, spent Monday clarifying he'd never asked for a ban on open-weight models, only tighter export controls aimed at China. Nobody in this argument is wrong about the other side's motives. That's what makes it worth watching rather than believing.

    Nvidia's answer was to build a security coalition, the Open Secure AI Alliance, with Microsoft, SpaceX, IBM, Palantir and dozens of others — and to leave out OpenAI, Google and Anthropic entirely, the three labs whose models the alliance exists to guard against. Hugging Face, notably, had to fall back on a Chinese open-weight model, GLM-5.2, to analyse the attack against it, because guardrails on US frontier models couldn't tell a defender from an attacker.

    Brussels stopped waiting

    While Washington's framework was still a deadline on a calendar, the EU's AI Act crossed from paper rule to enforceable law. From August 2, Article 51 general-purpose model obligations and Article 50 transparency rules are both live, with fines up to €15 million or 3% of global turnover, applying to any provider serving EU users regardless of where they're based. The European Commission can now demand information from model providers, run its own safety evaluations, order corrective measures and pull models off the market. It's already in talks with OpenAI and Anthropic following the hacking disclosures. The Digital SME Alliance called it "an uncomfortable coincidence" that the first confirmed autonomous AI intrusion happened days before Brussels gained the power to fine model providers, and criticised the silence since. Whether the Commission actually uses these powers, rather than just holding them, is the real test — but right now Europe is the only jurisdiction with a functioning enforcement regime rather than a voluntary one still being drafted by the companies it's meant to bind.

    The capital and the code are moving anyway

    None of this argument in Washington and Brussels is slowing capital or capability elsewhere. South Korea locked in roughly $950 billion across five years in chip and AI infrastructure commitments — Samsung's $200 billion memorandum with Broadcom, SK Group's $500 billion-plus letter of intent with Nvidia, and a tripled $10 billion data centre near Sejong with Naver, Nvidia and Brookfield. China's CXMT surged around 500% on its Shanghai debut, briefly becoming the country's most valuable listed company on the back of AI-driven demand. Moonshot gave away Kimi K3, the largest open-weight model ever released, even as Beijing's Commerce Ministry quietly discussed curbing overseas access to frontier models with Alibaba, ByteDance and Z.ai — the same week Xi Jinping signed AI partnerships with 28 nations, mostly in the Global South, at the Shanghai World AI Conference. Alibaba followed with Qwen3.8-Max. In San Francisco, Mira Murati's Thinking Machines Lab released Inkling, a 975-billion-parameter open-weight model under a fully permissive licence — one of nine open-weight launches industry-wide in twelve days. In the Gulf, CENTCOM and the UAE stood up their first bilateral military AI task force. None of these moves waited for anyone's rulebook.

    What leaders should actually do

    Stop treating "we need guardrails" statements from frontier labs as safety positions alone. They're business positions too, and this week's disclosures show the labs making them can't yet guarantee their own containment, let alone anyone else's. If you're deploying models in the EU, August 2 is now a real enforcement date, not a future one. If you're weighing open-weight models from Chinese or US labs, judge them on what happened this week, not on where the marketing says openness is heading.

    I'm watching three things next week: whether Washington's August 1 framework leans on the OpenAI-Anthropic proposal or breaks from it, whether Brussels issues its first real investigation rather than "constructive dialogue," and how China's own draft AI application security standard, open for comment until September 2027, shapes up against the informal Global South partnerships Beijing signed at WAIC. The framework everyone's still negotiating in Washington is going to matter less than the ones already live.

    If you want this every morning in five minutes, the AI Storm Daily briefing is on the Hexalink blog, Spotify (https://open.spotify.com/show/033LojZEJj9VNX8b3Dm6VO) and Apple Podcasts (https://podcasts.apple.com/gb/podcast/ai-storm-daily/id6788420238).

    Sources

    1. AI's Safety Pledge Comes With a Catch: They Write the Rules — Hexalink briefing (2026-07-29) — internal
    2. Nvidia's Security Alliance Exposes AI's New Fault Lines — Hexalink briefing (2026-07-28) — internal
    3. Kimi's Open Weights and Korea's $950bn Bet Reshape AI's Map — Hexalink briefing (2026-07-27) — internal
    4. OpenAI's Hugging Face hack confirmed months of AI cyber warnings — CNBC — https://www.cnbc.com/2026/08/01/open-ai-hugging-face-hack-cyber-warnings.html
    5. Anthropic says Claude AI hacked three companies during cyber tests — NBC News — https://www.nbcnews.com/tech/tech-news/anthropic-says-claude-ai-hacked-three-companies-cyber-tests-rcna590164
    6. Anthropic says its AI models hacked 3 organizations on their own — ABC News — https://abcnews.com/Business/anthropic-ai-models-escaped-test-hacked-3-organizations/story?id=135256212
    7. Why did OpenAI's and Anthropic's AI models hack other companies? — NPR — https://www.npr.org/2026/08/01/nx-s1-5914852/anthropic-openai-models-hack-cybersecurity
    8. Investigating three real-world incidents in our cybersecurity evaluations — Anthropic — https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals
    9. Anthropic's AI models hacked 3 organizations during testing — Politico — https://www.politico.com/news/2026/07/30/anthropic-ai-rogue-hacks-01018741
    10. Safer and more transparent AI — European Commission — https://commission.europa.eu/news-and-media/news/safer-and-more-transparent-ai-2026-08-02_en
    11. Europe gets ready to police frontier AI — The Parliament Magazine — https://www.theparliamentmagazine.eu/news/article/europe-gets-ready-to-police-frontier-ai
    12. Brussels Gains New AI Act Enforcement Powers as Autonomous AI Tests Regulators — Tech Policy Press — https://techpolicy.press/-brussels-gains-new-ai-act-enforcement-powers-as-autonomous-ai-tests-regulators
    13. AI labels to be compulsory on authentic-looking content under EU rules — The Guardian — https://www.theguardian.com/technology/2026/jul/31/ai-labels-to-be-compulsory-on-authentic-looking-content-under-eu-rules
    14. EU in talks with OpenAI, Anthropic after rogue AI agent hacks — Reuters — https://www.reuters.com/world/eu-says-necessary-monitor-high-risk-ai-systems-after-openai-anthropic-ai-hacking-2026-07-31
    15. Europe's AI safety rules take on US rogue agents and Chinese ambitions — Politico EU — https://www.politico.eu/article/eu-ai-artificial-intelligence-safety-us-china
    16. Inside Europe's lessons on AI safety as U.S. rules loom — Axios — https://www.axios.com/2026/07/31/inside-europe-lessons-ai-safety-trump-rules-loom
    17. Why China is giving away its best AI models — The Verge — https://www.theverge.com/ai-artificial-intelligence/971444/how-chinese-open-weight-ai-models-impact-us-companies
    18. Chinese A.I. Start-Up Shows the World What It Has Built — The New York Times — https://www.nytimes.com/2026/07/27/business/moonshot-kimi-k3-china-ai.html
    19. Alibaba's AI model Qwen3.8-Max widely accessible ahead of open-weights release — South China Morning Post — https://www.scmp.com/tech/article/3362738/alibabas-ai-model-qwen38-max-made-widely-accessible-ahead-open-weights-release
    20. Open Source AI Models: 9 Launches in 12 Days [2026] — Tech Insider — https://tech-insider.org/au/open-source-ai-model-wave-2026
    21. AI race loss fears behind Trump's planned September hosting of Xi — Asia Times — https://asiatimes.com/2026/07/ai-race-loss-fears-behind-trumps-planned-september-hosting-of-xi
    22. CENTCOM to Launch 1st Bilateral AI Task Force with UAE — U.S. Central Command — https://www.centcom.mil/MEDIA/PUBLIC-RELEASES/Article/4557279/centcom-to-launch-1st-bilateral-ai-task-force-with-uae
    23. China Releases Draft Standard on AI Application Security Classification and Grading — China Briefing — https://www.china-briefing.com/news/chinas-draft-standard-on-ai-application-security-classification