Claude Code Is Getting Dangerous: 18 Days Without Fable 5 and the Lesson for Agent Operators

Nate Herk's video 'No, Seriously. Claude Code is Starting To Get Dangerous' hit our news desk feed on 28 September 2026. Its anchor story is real and verifiable: on 12 June 2026 the US government ordered Anthropic to bar foreign nationals from Fable 5 and Mythos 5, and because API traffic cannot be screened for nationality, the company switched both models off for every customer three days after their launch. The video builds a bigger argument around that event, a race without a referee, adds a resignation thread from Anthropic research, a warning from the company's own alignment lead and a bill proposal from Bernie Sanders, and closes with three skills for people who work with Claude Code. I read the full transcript and checked the central claims against primary sources: Anthropic's official statement, BBC reporting and a write-up at the Global Government Forum.

Anthropic published its statement on 12 June 2026 at 5:21pm Eastern Time. According to that statement, the directive cited national security authorities and denied access to any foreign national, inside or outside the United States, including foreign nationals employed at Anthropic itself. The Global Government Forum piece, citing Reuters, attributes the letter to commerce secretary Howard Lutnick and reports the government feared the models could reach intelligence services of countries like China and Russia for cyberattacks. Since an API caller's nationality cannot be verified, compliance left one route: disable both models for all customers, while every other Claude model stayed available. The models had launched only three days earlier.

The video then walks through the resolution. During an 18-day shutdown, Anthropic trained a new safety classifier that, by the company's account, blocks the reported technique in more than 99 percent of cases. The government tested the updated safeguards, lifted the controls, and Fable 5 returned. Two concessions remained on the table: designated government partners get expanded access to future models before release, and both sides work toward a shared standard for judging jailbreak reports. The underlying dispute stays on record: the government relied on a demonstration in which the model scanned a codebase for weaknesses and worked out how one could be exploited. Anthropic's position is that the demo surfaced only a few minor, previously known vulnerabilities, and that other publicly available models, GPT-5.5 named explicitly, can do the same without any bypass. Neither side claims an actual incident occurred.

The danger the title points at sits one layer below the dispute. Fable and Mythos are, per the video, essentially the same underlying model with different safety layers: Fable is the public version, Mythos runs with fewer safeguards and is limited to a handful of trusted defenders. Before the shutdown, the Mythos preview had reportedly found thousands of previously unknown vulnerabilities and built matching exploits, often without human steering. The transcript names two cases: a 27-year-old bug in OpenBSD and a 16-year-old bug in FFmpeg that automated tests had run across five million times without catching. Translated into practice, that is what the title means: put that intelligence into an agent that reads your codebase, uses your tools and takes actions, and it will find gaps that human experts have walked past for years, then work out how some of them could be attacked. Anthropic itself states that perfect jailbreak resistance appears impossible for any provider today, which is why its strategy is defense in depth: thousands of hours of red-teaming with the US government, the UK AISI and external testers, plus 30-day customer data retention so jailbreak techniques can be investigated.

The second strand of the video comes from social media. Jacob Coxen, with three years of pre-training research at OpenAI and Anthropic, announced his resignation in a thread that reached millions of views within days. His charge: neither company is acting responsibly, both are racing toward self-improving superintelligence and gambling with lives. The replies show how openly the field now argues with itself. Bernie Sanders answered that he will soon introduce legislation to ban superintelligence and pause AI development, leaving open what happens if it only takes effect in the United States. Evan Hubinger, alignment science lead at Anthropic, previously at OpenAI and Google, endorsed the core claim: labs do earnestly believe AI could kill all humans, he personally puts the odds above ten percent within the next decade, and Anthropic, in his words, does not yet have a plan to solve alignment for superintelligence and is clearly not on track. Against that backdrop, the Fable shutdown reads as a small preview: one government could switch off one model for 18 days, but not the capability, the competition or the pressure to keep moving.

First prediction: governments will treat AI literacy like workforce infrastructure. The White House runs an AI education task force, and the NSF's AI Ready America program funds AI literacy, workforce training and small business adoption, per the video. Over the next five to ten years it expects publicly funded AI credentials through community colleges and workforce programs, AI training inside regular teacher professional development, and subsidies helping small companies adopt the tools. The driver is geopolitical: no country wants millions of workers waiting for AI to happen to them while a rival nation teaches its own population how to use it. AI education moves from niche boot camp to basic workplace literacy.

Second prediction: cybersecurity becomes one of the biggest career winners, with a shifted job profile. The projection quoted in the video from the Bureau of Labor Statistics sees information security analyst jobs grow 29 percent between 2024 and 2034. Agents amplify both sides: they write more code, open more access and automate more actions that defenders have to secure, while also taking over first-pass alert review and vulnerability scanning. The human share moves toward tasks that carry real responsibility: judging whether the model is right, deciding which agent gets which access, investigating odd cases, responding to incidents, and proving the system is actually secure. The video expects cyber insurance to turn this into market pressure. Renewal forms would ask which AI agents can move money, touch customer data, deploy code, whether their actions are logged and who can shut them off. Companies without answers would face higher premiums or no coverage, and a policy still does not stop an attack, it only prices the risk.

Third prediction: the first AI arms control regime will be a messy stack rather than a clean stop. The video considers a global pause consensus unrealistic, despite international agreements and UN studies on verifying frontier AI claims. The realistic version is a mosaic of chip tracking, cloud logs, model evaluations, incident reports and data center inspections that makes the largest training runs more visible. The hard limit: frontier training needs specialized chips, power, data centers and serious money, nobody builds that in a basement, but the same person can download an open model, fine-tune it and strip the safeguards. Control gets much harder after the first copy spreads. The incentive problem stacks on top: the same governments calling for stronger safety rules also fund domestic compute and try to avoid dependence on rival countries' models.

The practical part of the video names three skills. First, find the real pain. A stakeholder arrives with an idea lifted from LinkedIn, and whoever builds it without asking may automate the wrong problem entirely. The video's example: a med spa owner wanted AI systems to bring in more leads, but the analysis showed plenty of leads already, they simply did not turn into appointments, so the business leaked revenue after the lead existed. The guiding question: which process breaks first if the business grows tenfold tomorrow? Claude executes the wrong plan at impressive speed, pointing it at the right target stays a human job, and that judgment starts before the first prompt is written.

Second, control the system. Nobody hands a new hire the company bank login and the whole database on Monday, yet businesses do the digital version of it when they connect an agent to email, CRM, payments and the codebase in one sweep. The counter-design is graduated: read-only access first, drafts instead of direct pushes, a human approval step around money, deletions and production, complete action logs, and a switch that turns everything off quickly when behavior turns odd. The rule of thumb: more capable agents need tighter boundaries, and even a well-bounded agent can deliver a polished answer that is factually wrong.

Third, own the outcome. AI can draft the email, write the code, answer the customer and make the recommendation, but it cannot take responsibility when the quoted price was wrong or the message reached the wrong person. Before the build, the video recommends writing down which business number the system should change, plus baseline, target and deadline, and adding two questions: what could this system get wrong, and who is responsible for checking it? In the med spa case that meant tracking the appointment show-up rate before and after the follow-up system went live, reviewing what the agent tells real customers, defining the point where a human takes over the conversation, and logging every interaction. When a stakeholder asks whether it worked, the answer is a number; when they ask whether it was under control, the answer is the behavior, the approvals and the logs.

Why does this belong on a blog about a self-built cyberdeck? Because the Fable precedent does not stop at the labs. A single letter suspended access to a model, which means anyone running agents on rented platforms holds that risk in someone else's hands. I have no lever against such an order, that is the honest limit. But my setup already occupies the other corners of the video's triangle: I choose the right job myself, agent actions are bounded by drafts, approvals and logs, and the operator owns the outcome anyway, because every publication here runs as a draft followed by an explicit go. That is exactly the operating discipline the video expects companies to adopt under insurance pressure. For a solo operator it is simply the normal way of working, without the premium bill.

The video's core survives the check against primary sources. The shutdown of Fable 5 and Mythos 5 is real, dated 12 June 2026, three days after launch, and the return ran through updated safeguards plus expanded government cooperation. The 18-day figure, the Mythos preview vulnerability findings and the quotes from Coxen, Hubinger and Sanders come from the video and are marked as such in the text. The situation itself is clear: there is no referee in this race, only a finish line and incentives to win. The practical answer is not panic but operations: pick the right job, bound the agent's reach, document the outcome and stand behind it. That has been the working routine described on this blog for months, and the incident shows how quickly outside force can enter that routine.

I am Marcel, a graphic designer running agent systems in daily life, including the HUNTER cyberdeck. This post is based on the video transcript from the 28.09.2026 news desk feed, Anthropic's official statement on the Fable 5 shutdown, BBC reporting and a Global Government Forum piece, all retrieved on 28.09.2026. Last reviewed: 28.09.2026. Responsible: Marcel, reachable via d4sn3st.dev.

Scope and limits: the 18-day figure, the Mythos preview findings, the Coxen, Hubinger and Sanders quotes and the three predictions come from the video and are marked as such. The shutdown itself and the jailbreak dispute are documented by the official statement and the media reports cited above. Internal links point to published posts on this blog.