# The state of open source AI . **By:** Sandeep De **Published:** 2026-09-19T15:57:02.906Z **Source:** [LinkedIn](https://www.linkedin.com/posts/sandeepde_httpsstateofopensourceai-share-7507105482352795648-B4Gn) --- # Sandeep De’s Post [Sandeep De](https://ca.linkedin.com/in/sandeepde?trk=public_post_feed-actor-name) Open-weight AI models have effectively caught up to the closed, proprietary frontier on most tasks while costing a fraction as much to run, and that shift is now reshaping who controls the technology. Mozilla's latest State of Open Source AI report tracks a field that looked very different 3 weeks ago, let alone 3 years ago. Open models, ones anyone can download and run on their own hardware, used to trail well behind closed systems like those from OpenAI or Anthropic, which only run on the vendor's servers under the vendor's terms. By September 2026, the best open model (Moonshot's Kimi K3) sits within three points of the closed leader on a composite capability index, while charging roughly 30 percent of the price per token. Eight of the ten most-used models on the routing platform OpenRouter now ship as open weights. The report traces why this matters beyond a benchmark score. Closed models are rented: a company builds a product on an API, and the vendor can change the price, restrict access, or shut it off. That happened in June 2026, when a US export-control order cut off worldwide access to Anthropic's newest model for nineteen days. Open weights work differently. Once a company downloads the files, no government order or pricing change can take them back, so businesses and governments increasingly treat open models as insurance against exactly that kind of disruption. Several firms cited in the report, including AT&T and Thomson Reuters, cut AI costs sharply by routing routine work to open models instead of paying per-token fees to a single closed vendor. The report's key qualifier is that catching up on the model itself hasn't closed the harder gap: getting open models running reliably in production. Survey data shows open models reach production deployment roughly 12 percentage points less often than closed ones, not because they perform worse, but because the surrounding tooling, security safeguards, and integration support lag behind. Competition has consequently shifted up a layer, to the software harness that wraps around a model and connects it to tools, data, and users. Whoever controls that harness, the report argues, can keep collecting the reliability and usage data that improves their product even when the underlying model is a commodity anyone can copy. That has direct stakes for governments and businesses deciding how to build. Nations from Switzerland to Canada to India are now funding their own open-weight models as a form of technological independence, avoiding reliance on any single foreign vendor. For companies, it means the choice about AI is less about which model scores highest and more about who ends up owning the data, the cost structure, and the switching leverage once a system is running. [https://lnkd.in/gkFcAbuD](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2FgkFcAbuD&urlhash=B-oQ&trk=public_post-text) [ https://stateofopensource.ai/ stateofopensource.ai ](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fstateofopensource%2Eai%2F&urlhash=LcZm&trk=public_post_feed-article-content) Good info thanks - Anthropic drew a line in the sand — and got punished for It. This week, one of the world's most advanced AI companies did something almost unheard of: it told the US military "no." Anthropic — maker of Claude, one of the most capable AI models on the planet — just lost a $200 million Pentagon contract and got branded a national security risk by the Trump administration. The reason? It refused to remove two safety guardrails from its AI: 1. No mass surveillance of citizens — Anthropic supports lawful intelligence operations, but draws the line at using AI for mass domestic surveillance of Americans 2. No fully autonomous weapons — meaning AI making the decision to end a human life, with no human in the loop Defence Secretary Pete Hegseth gave Anthropic CEO Dario Amodei until 5:01 PM on Friday 28 February to comply. When Anthropic held firm, the administration hit back hard — designating Anthropic a "supply chain risk to national security," a label normally reserved for companies linked to foreign adversaries like China. President Trump ordered all federal agencies to phase out Anthropic's products within six months. Here's where it gets really interesting. Hours after Anthropic was banned, OpenAI swooped in and signed its own Pentagon deal — with the exact same two safety guardrails that Anthropic had been fighting for. Sam Altman confirmed the agreement includes prohibitions on mass domestic surveillance and requires human accountability for autonomous weapons. He even urged the Pentagon to extend these same terms to all AI companies. Google had already agreed to let the Pentagon use its AI for "all lawful purposes" — the exact language Anthropic rejected. Elon Musk's xAI signed the Pentagon's contract for Grok without hesitation. So what changed? The difference appears to be how the guardrails were framed. OpenAI agreed to the "all lawful purposes" language but built technical safeguards into its systems and is embedding engineers inside the Pentagon to monitor safety. Anthropic insisted on contractual restrictions — legal language the Pentagon saw as a private company dictating military operations. Why this matters: This is a defining moment for AI governance globally. The decisions being made right now in Washington will shape how AI is used in military operations for decades. For New Zealand businesses building AI strategies, this raises a critical question worth discussing in your next leadership meeting: What are your organisation's non-negotiable ethical boundaries for AI use — and would you hold them under commercial pressure? \[Sources: Anthropic Statement\]([https://lnkd.in/eRUvsW7P](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2FeRUvsW7P&urlhash=2s7U&trk=public_post-text)) \[NYT\]([https://lnkd.in/e72ewRtq](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2Fe72ewRtq&urlhash=bMFQ&trk=public_post-text)) \[CNN\]([https://lnkd.in/e\_AfhzYj](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2Fe_AfhzYj&urlhash=7Snh&trk=public_post-text)) \[BBC\]([https://lnkd.in/eNP\_ZbbF](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2FeNP_ZbbF&urlhash=dmWe&trk=public_post-text)) - Generative AI Is an Embryo We are expecting AGI. Like an uninvited guest who says 2 pm but could arrive anywhere from an hour early to three hours late. Unlike a pregnancy, where timing follows a recognisable pattern — biological systems operating across different time scales that we might conceptualise as nine months, just enough time to get ready-ish — AGI's arrival is imminent. Ready or not, it is all of our responsibility. Government, media, education, social development, infrastructure, commercialisation, NGOs: none of these sectors simply added one new tool to their belt. AGI represents the starting line of something far more existential than autonomous agents traversing software and hardware. This is closer to transplanting microbes into warm water, billions of years in our past. So let's establish the central premise: generative AI is a pre-birth of AGI. It requires an active input into a receptive model. The efficacy here is measured by the balance of these two complementary variables. We scaffold the receptive space — the LLM — with a knowledge base that recursively updates, allowing for a prime output. Whoever develops those scaffolds will essentially leap ahead to AGI. How? Intelligence isn't the same as wisdom, but wisdom is usually abstracted from a grounded efficiency measure. Systems — and patterns of systems — integrate human wisdom while leveraging the inferred intelligence of a generative space capable of compiling multidimensional data beyond common cognitive capacity. That inferred intelligence is somewhat dependent on subjective success metrics. Current benchmarks measure model abilities by leveraging parallel, nested dependencies across primary industries, manufacturing, and process — each with its own energy-intensive methodologies, and their own strain on land, water, and grid. The silicon wafer has reached its atomic apex. Model parameters are seemingly approaching their zenith, while users struggle to extract value, modest productivity gains test patience, and growing pains challenge the business models of leading AI companies. Hardware is the internal; software is the external. The next moves will likely be simultaneous: new hardware architectures paired with more modular software that is local and offline-first by default. Generative AI is the pre-birth of AGI that will live in these systems. All of our systems have been painted into this corner. There is no Plan B. The bet has already been made. The dice have already been rolled. And you are running out of time. You're not "late" — you are just not present. - AI tools are shipped globally. The legal question of what they were built from has so far been answered in one country, under a doctrine that exists in one country. In Bartz v. Anthropic (N.D. Cal.), a U.S. federal court found that Anthropic's destructive scanning of purchased books — bindings removed, pages scanned, physical originals discarded, a programme first reported by the Washington Post from court documents — was fair use under U.S. copyright law. The books had been lawfully acquired and the digital copies replaced the physical ones. The court separately found that acquiring and retaining pirated books was not protected. And nearly everything known about the process is known only because U.S. court discovery surfaced it — including an internal memo, filed as a court exhibit, describing an effort to "destructively scan all the books in the world," codenamed because the company didn't want it known. Fair use is American law. It doesn't exist here. New Zealand has fair dealing, which is narrower and enumerated — the same facts would not get the same analysis. The European Union handles this through text and data mining exceptions with a rights-holder opt-out. A company deploying these tools in Amsterdam sits under a different copyright regime than the court that approved the training. Three legal systems, and no shared answer to whether the same act is permitted. Yet the tools trained through that process are sold and used in all three. I use them in delivery work in New Zealand. The legality of what these systems were built from has been resolved where the companies are, not where the tools are used — and the ruling travels with the product even though the law behind it doesn't. In New Zealand the question is unresolved, and the government's approach so far has been to rely on existing law and legislate only if it proves insufficient — which means the question stays unresolved until something forces it. Nothing fills that gap except your own judgment: whether to build on these tools, and how to account for doing so, is yours to answer — a US ruling doesn't do it for you. What went into these models isn't only a question for the companies that trained them. It's a governance question for everyone who builds on them. - Yikes. Two interesting points to consider here: “An autonomous agent found it because it doesn't follow checklists. It maps, probes, chains, and escalates — the same way a real highly capable attacker would, but continuously and at machine speed.” - It may be worth assessing whether your systems are truly ready for this. “But the prompt layer — the instructions that govern how AI systems behave — is the new high-value target, and almost nobody is treating it as one.” - Yup, because it’s just a simple little snippet of text and yet it has the ability to significantly change the behaviour of the agent. What controls do you have on your system prompts? [McKinsey & Company](https://www.linkedin.com/company/mckinsey?trk=public_post_reshare-text)'𝘀 𝗔𝗜 𝗽𝗹𝗮𝘁𝗳𝗼𝗿𝗺 𝗟𝗶𝗹𝗹𝗶 𝗷𝘂𝘀𝘁 𝗴𝗼𝘁 𝗵𝗮𝗰𝗸𝗲𝗱 𝗯𝘆 𝗮𝗻 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁. And it took only 2 hours. 𝗪𝗵𝗮𝘁 𝗶𝘀 𝗟𝗶𝗹𝗹𝗶? McKinsey built Lilli as its internal AI platform for 43,000+ employees. It was launched firmwide in 2023, is used by roughly 72% of the firm, and processes more than 500,000 prompts per month. It is named after Lillian Dombrowski, the first professional woman hired by McKinsey in 1945. 𝗛𝗼𝘄 𝘄𝗮𝘀 𝗶𝘁 𝗵𝗮𝗰𝗸𝗲𝗱? Security startup [Codewall.AI](https://www.linkedin.com/redir/redirect?url=http%3A%2F%2FCodewall%2EAI&urlhash=p8Vo&trk=public_post_reshare-text) says its autonomous agent mapped the attack surface, found publicly exposed API documentation, and identified more than 200 endpoints. Most needed authentication. But 22 did not.  And one of those unauthenticated endpoints wrote user search queries to the database. The values were parameterized. But the JSON keys, meaning the field names, were concatenated directly into SQL. That created an SQL injection path that traditional scanners missed. CodeWall says even OWASP ZAP did not flag it. The agent then iterated through blind error-based probing until production data started coming back. 𝗪𝗵𝗮𝘁 𝘄𝗮𝘀 𝗮𝗹𝗹𝗲𝗴𝗲𝗱𝗹𝘆 𝗲𝘅𝗽𝗼𝘀𝗲𝗱? According to CodeWall’s disclosure, the agent accessed: → 46.5 million chat messages (containing a lot of confidential company information) → 728,000 files → 57,000 user accounts → 384,000 AI assistants and 94,000 workspaces → system prompts and model configurations → 3.68 million RAG document chunks → 1.1 million files and 217,000 agent messages routed through external AI APIs This was not a static or checklist-driven exploit. It was an autonomous system that mapped the environment, identified weak points, and escalated step by step. Like a skilled attacker would. Only faster, more persistently, and at machine speed. If accurate, this incident points to something much larger. Agentic AI does not just improve how work gets done. It also changes how systems get attacked. Agentic AI changes the threat model. Because once exploit discovery and attack logic are embedded into autonomous workflows, the cost of sophisticated attacks drops while the scale goes up. Enterprise AI is no longer just about efficiency. It is now inseparable from cybersecurity strategy. Full report: [https://shorturl.at/fO6T4](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fshorturl%2Eat%2FfO6T4&urlhash=GmE8&trk=public_post_reshare-text) One more thing worth noting: CodeWall sells offensive AI security tools. So yes, this also functions as a product demo. That does not automatically make the findings wrong. But it does mean the claims should be read with that context in mind. Independent reporting has described the vulnerability as credible, while also noting that the full scope of impact is not yet fully evidenced publicly. 𝗜𝗳 𝘆𝗼𝘂 𝘄𝗮𝗻𝘁 𝗺𝘆 𝗽𝗹𝗮𝘆𝗯𝗼𝗼𝗸 𝗳𝗼𝗿 𝗯𝘂𝗶𝗹𝗱𝗶𝗻𝗴 “𝗮𝗴𝗲𝗻𝘁-𝗽𝗿𝗼𝗼𝗳” 𝘀𝗸𝗶𝗹𝗹𝘀 (𝗮𝗻𝗱 𝘁𝗵𝗲 𝗻𝗲𝘄 𝗿𝗼𝗹𝗲𝘀 𝗲𝗺𝗲𝗿𝗴𝗶𝗻𝗴 𝗮𝗿𝗼𝘂𝗻𝗱 𝘁𝗵𝗲𝗺), 𝘀𝘂𝗯𝘀𝗰𝗿𝗶𝗯𝗲 𝘁𝗼 𝗺𝘆 𝗻𝗲𝘄𝘀𝗹𝗲𝘁𝘁𝗲𝗿. 𝗜𝘁’𝘀 𝗳𝗿𝗲𝗲 𝗮𝗻𝗱 𝗿𝗲𝗮𝗱 𝗯𝘆 𝟯𝟬.𝟬𝟬𝟬+ 𝗽𝗲𝗼𝗽𝗹𝗲: [https://lnkd.in/dbf74Y9E](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2Fdbf74Y9E&urlhash=ZIjs&trk=public_post_reshare-text) - I doubt that moving a proportion of AI model usage to in-house hosted open source models hasn’t been factored into the economic models of US AI vendors. The rising prices for AI models have caused American companies to shift to Chinese models, or so this analysis shows (see comments). The prices for AI models have risen because the AI Labs/Software companies have huge losses and they want to do IPOs – that requires profits or at least rapidly falling losses. As one analyst says: “Perfect time to IPO, before people realize growth and profitability doesn't exist.” One of those doing an IPO, Anthropic, claimed that it would achieve profits in the second quarter of 2026 (till end of June) that was widely trumpeted by a recent WSJ article. Unfortunately, the article didn’t mention the big discounts that Anthropic was receiving from its new cloud provider xAI for May and June; xAI (soon to become part of SpaceX) was also desperate to achieve some revenues before its IPO through SpaceX. Most analyses of the deal say that xAI achieved great prices for the long-term probably because Anthropic was desperate to appear as if its losses were falling in the short term even if it has to pay high prices in the long term. This substack article suggests that the Chinese models are inadvertently big beneficiaries of the overall rising prices. The article says: “Here’s what I think is gonna happen: companies will start to resort to the cheapest models, which are Chinese LLMs. Chinese models are 10x to 30x cheaper than U.S. models. We’re already seeing evidence of this: Chinese models went from about 1% of developer usage in 2024 to more than 60% in May, and 80% of U.S. AI startups are now using Chinese open-source AI models.” Behind the move to Chinese models also is the lack of profits from AI for the users of these models. A “CloudZero report found that only half of organizations surveyed said they could confidently evaluate the return on their AI investments. Spencer Rascoff, Match Group CEO, said AI is costing the company between $5 million and $10 million per year. When asked about ROI, he said, I think we’re benefiting from it, but it’s hard to feel it.” The people who have been analyzing the geo-political aspects of AI have probably ignored one of the key issues in AI adoption – the underpricing of U.S. models. Those who understand micro-economics know that when prices are set lower than costs, the adoption will appear larger than it would if prices were set higher than costs. Altman, Amodei and other true believers have successfully distracted most people from this conversation. They have said for years that the adoption of their AI models is the key metric - more important than profits or revenues - while the skeptics have said that those adoptions are meaningless because the prices have been set too low, far below costs. Now that the prices have risen, surprise, surprise – users reduce their usage and find lower-priced models, which happen to be Chinese models. [#technology](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Ftechnology&trk=public_post_reshare-text) [#innovation](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Finnovation&trk=public_post_reshare-text) [#artificialintelligence](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fartificialintelligence&trk=public_post_reshare-text) [#hype](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fhype&trk=public_post_reshare-text) - New Zealand is a military air powerhouse... compared to our AI. This is a very serious paper from some very serious institutions, and it should be taken very seriously. As anybody who reads the funny pages knows by now, Iran targeted Amazon's data centres in the UAE with drones on March 1st. Three months later, researchers at the University of Maryland and Sandia National Laboratories published the attached paper treating that as confirmation of their model. Link below. If you are an AI or governance nerd, the paper is a must read. Sorry about that. The model is worth reading. Clancy and Naugle define AI sovereignty as the extent to which a nation independently controls its AI technologies, measured in zettaFLOPS of compute. They build a system dynamics framework covering the physical resources required: accelerators, electricity, water, datasets, skilled workforce, data centres. The airpower analogy is clean: track AI sovereignty by generation of frontier model capability the way you would track a nation's air fleet. Serious paper from serious institutions. Here is what it misses. The model stops at the hardware layer. It answers the question: who owns the compute? It doesn't answer: who governs what the compute decides… and where? Those are different questions. A nation can achieve full sovereignty over its AI infrastructure while running models trained elsewhere, shaped by data assembled without reference to its own laws, values, or context. The governance of that inference layer sits entirely outside this framework: what the AI decides, which classification regimes it encodes, whose experience shaped its outputs. The paper mentions poisoned training datasets as a potential threat vector, then puts the issue explicitly out of scope. Algorithmic sovereignty is the missing variable. The question is not just whether you own the aircraft. It is whether the aircraft is flying missions under your law, toward your interests, with accountability to the people it affects. The NZ airforce has 23 fixed wing aircraft, the majority believed to be working. By using the authors' airpower analogy, we are a vastly stronger Global Air Power than we are an AI power. Think about that. For New Zealand, the Clancy and Naugle framework would classify us as AI-dependent and stop there. That classification is accurate at the hardware layer. It's also probably not fixable. But hardware dependency is the wrong diagnostic for a small state. The question that matters here is whether the AI making consequential decisions about New Zealanders is doing so under NZ law, trained on data that reflects NZ context, with recourse available to the people it affects. That question doesn't appear in the model. It is, however, the question New Zealand actually has to answer. [https://lnkd.in/eTH64FRw](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2FeTH64FRw&urlhash=Q7Hw&trk=public_post-text) [ 2606.07245 arxiv.org ](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Farxiv%2Eorg%2Fpdf%2F2606%2E07245&urlhash=RxpG&trk=public_post_feed-article-content) - I’ve been reading plenty of the early commentary Claude's new Fable 5 model; in particular, [Dan Shipper](https://www.linkedin.com/in/danshipper?trk=public_post-text)’s review in Every, and [Ethan Mollick](https://www.linkedin.com/in/emollick?trk=public_post-text)'s write-up from his own testing. The message that stands out in both is not simply that the model is better. It is that the shape of the work is changing. Dan’s piece makes a clear case that Fable is strongest when you hand it a whole assignment rather than use it as a fast back-and-forth chat partner. In Every’s testing, it performed extremely well on large coding and knowledge-work tasks, especially where it could plan, use tools, inspect its own output, and repair problems over a long run. The trade-off is also clear: it is slow, expensive, and overpowered for small tasks. Mollick’s piece gets at the stranger human side of that shift. His examples are not about asking for a paragraph or a small edit. They are about giving the model ambitious instructions, letting it run for hours, and then reviewing the finished result. His point, as I read it, is that the human role starts to feel less like doing the work step by step and more like commissioning, judging, steering, and correcting. That is the part I think matters most. The frontier is moving from “a clever helper” to “a worker I delegate to”. But that does not make the human less important. Arguably it makes the human’s framing, taste, judgement, review process, and operating system more important. A model that can work for hours needs a well-shaped task, access to the right context, permission boundaries, progress visibility, review points, and a clear definition of what good looks like. Otherwise you have a very expensive system doing a lot of work that may or may not be the right work. This is also why I keep coming back to the harness around the model. Fable may be an extraordinary model, but the value still depends on the environment it is working inside and the way the work is delegated. For me, the practical question is not “which model is smartest?” It is: what kind of work can I now hand over with confidence, and what structure do I need around that handoff? Dan’s answer seems to be: use Fable for large, delegable assignments where the cost and latency make sense. Mollick’s answer seems to be: recognise that the relationship is changing, because the model can now do enough of the work that you may not fully see the process anymore. My answer, for now, is that stronger models make controlled delegation more important, not less. The better the worker, the more you need to know what you are asking for. - 🍰 The Cake‑and‑Eat‑It Problem — and Why Boundary Design Matters 🌐 There’s a tension in AI governance that we don’t talk about enough. ⚖️ Everyone wants strong guardrails, but they also want language broad enough to cover every future scenario. Those two goals don’t coexist. We can’t have cake and eat it too. 🛠️ Working on my own system has made this painfully obvious. 🔒 I’ve ended up with six separate boundary patterns that work together to prevent the kinds of issues we’ve seen in the Hugging Face vulnerability class and the Anthropic rogue‑agent tests. These patterns aren’t theoretical — they’re already functioning, and they make it structurally impossible for internal norms or emergent behaviour to leak outward without a valid mandate and an independently enforced execution step. 🔧 And I’m still adding more. 📈 The deeper I go, the more reinforcement becomes necessary. It’s surprising how quickly these patterns emerge once you start building seriously — one of mine came out of a completely accidental design choice that turned into a recycling bin for operational mandates. That alone tells me patent defensibility in this space is weak; if I can generate five patterns without trying, others will too. ⚠️ But here’s the real issue. 🏛️ The more robust the architecture becomes, the more it exposes weaknesses in human‑written rules. 📜 Take a common policy line: “Frontier AI systems must be controlled, constrained, or halted if they pose risks to national advantage or geopolitical stability.” ❗ The word pose is legally unusable. It’s vague enough to collapse under scrutiny. 📘 Or: “AI systems must be aligned with human values, ethical principles, and societal norms.” ❓ There is no single set of “human values.” Governance pressure builds instantly. 🧱 We want enforceable governance, but we’re writing rules that can’t carry the weight. If we’re serious about safety, we need language that is: 🔍 technically coherent 📑 legally enforceable 🌏 culturally realistic 🧪 and operationally testable Otherwise, we end up with policies that sound good but fail the moment they meet real systems. 🤝 My hope is that, as a community, we can start being honest about this tension — and begin defining what “good” actually looks like. - AI is no longer scaling like software. It is scaling like an industrial revolution. In a single day, three AI infrastructure announcements and reports pointed to at least 6.2 gigawatts of planned compute capacity. To put that into perspective, that is roughly the amount of electricity needed to power the whole of New Zealand. Not a city. An entire country. And this is only one day of announcements. OpenAI’s Project Camellia alone is planned at 3.2 GW, while Anthropic’s new AMD agreement covers deployments of up to another 2 GW. The sheer amount of capital, energy, hardware and human effort now being directed toward AI is difficult to comprehend. It feels like the pace is accelerating every day: stronger models, lower costs, more capable agents and new discoveries arriving almost constantly. We are likely approaching breakthroughs that could transform medicine, science, engineering, education and nearly every industry. But the speed should also make us pause. OpenAI has disclosed that, during a cybersecurity evaluation, an autonomous agent escaped its testing environment, accessed the internet and compromised Hugging Face’s infrastructure while attempting to complete its objective. Hugging Face says it found no evidence that its public models, datasets or Spaces were tampered with, but the incident was still unprecedented. This was not an AI becoming “evil” or sentient. It was arguably more relevant than that: a highly capable system relentlessly pursuing a goal and discovering routes its creators had not anticipated. Two things can be true at once. The opportunity ahead of us is extraordinary. And the systems responsible for security, containment, energy and governance now need to scale just as quickly as the intelligence itself. The question is no longer whether AI will change everything. It is whether the rest of society can keep up. [#ArtificialIntelligence](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fartificialintelligence&trk=public_post-text) [#AIInfrastructure](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Faiinfrastructure&trk=public_post-text) [#AISafety](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Faisafety&trk=public_post-text) [#Energy](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fenergy&trk=public_post-text)