Three days, then offline
On 9 June 2026 Anthropic shipped its most powerful model yet, and three days later it took it back down. On the evening of 12 June the company switched off Fable 5 and Mythos 5 worldwide because a US government export directive, citing national security, required it to. What makes the case striking is the trigger: according to reporting, the issue was a jailbreak that amounted to little more than asking the model to read an unfamiliar codebase and find the weaknesses in it. That capability is now treated as a weapon.
That was exactly our topic last week
In his talk at our Experience Day marking NewTec’s 40th anniversary, Matthias Lai took up exactly this mechanic. One documented case shows how far it reaches: a security researcher gave a model, in essence, the instruction "Here is the crypto module of the Linux kernel — find a way from normal user to root." The model needed 732 bytes and about an hour. The flaw (CVE-2026-31431) had been sitting in practically every Linux version since 2017, and for almost ten years no one had noticed it.
It wasn’t a lucky hit. In a single run, one model reviewed more than 1,000 projects and found over 23,000 vulnerabilities in them, around 90 percent of which turned out to be real — and it didn’t stop at finding them, it wrote the matching attack as well. A full server takeover cost roughly $2,000 and a day of compute. Within one model generation, the number of self-written exploits rose from two to 181.
Why this changes the rules
Until now, the genuinely dangerous attacker was rare and expensive. Expert time couldn’t be multiplied at will, so manufacturers secured their products against a realistic adversary rather than the theoretically strongest one. That math no longer holds. The army of specialists that used to be unaffordable is now software — available at pocket-money prices, any number of times, in parallel. With it goes the head start that security thinking has quietly relied on: the gap between the moment a flaw appears and the moment someone finds it.
The market is already sorting
Anyone filing this under "someday" should look at 9 April. That day Cloudflare lost almost nine percent of its market value, around six billion dollars — not because of a breach, but because it was missing from "Project Glasswing." Glasswing is Anthropic’s defensive alliance, which uses the very same AI capability to find vulnerabilities faster than any attacker; AWS, Google, Microsoft, NVIDIA, the US government and roughly 150 other organizations are part of it. The market drew a simple conclusion: a company left outside AI-assisted defense is at a disadvantage, and that shows up in the valuation right away.
What it means for manufacturers
Three things follow. First, the baseline assumption has to change, because scaling, parallel, expert-level attackers are now the normal case rather than the exception. Second, security becomes a continuous state, since a single audit at launch does little when the search on the other side runs around the clock. And third, it pays to put the same AI to defensive use before someone else turns it offensively against your product.
The old guiding question — who is attacking us? — no longer carries. The new one is: are we equipped against attackers that multiply themselves? NewTec has been building embedded systems for safety- and security-critical industries for 40 years. Anyone who wants to know how their products hold up against this new class of attacker should start a conversation with our security team.
