Saturday, June 13
Saturday, June 13
Anthropic's Best Models Lasted Four Days Before the Government Pulled the Plug
Anthropic killed access to Fable 5 and Mythos 5 for all users Friday evening after Commerce Secretary Howard Lutnick issued an export control directive triggered by an unnamed competitor claiming it jailbroke Mythos. The models launched Monday. They were dark by Friday. Anthropic says the government offered only "verbal evidence" of a "narrow, non-universal jailbreak" that basically asks the model to read code and fix flaws. Which is what coding models do. Enterprise customers who started building on Monday watched their tools vanish by government order. The political backstory is generating fierce debate about whether this is national security or something uglier.
Anthropic's Best Models Lasted Four Days Before the Government Pulled the Plug
Anthropic killed access to Fable 5 and Mythos 5 for all users Friday evening after Commerce Secretary Howard Lutnick issued an export control directive triggered by an unnamed competitor claiming it jailbroke Mythos. The models launched Monday. They were dark by Friday. Anthropic says the government offered only "verbal evidence" of a "narrow, non-universal jailbreak" that basically asks the model to read code and fix flaws. Which is what coding models do. Enterprise customers who started building on Monday watched their tools vanish by government order. The political backstory is generating fierce debate about whether this is national security or something uglier.
AI Security Watch
The tools developers trust most are becoming the tools attackers target first. That pattern is old. What's new is how the AI layer collapses the distance between "trusted input" and "executed command."
- The average developer now runs 3-5 AI-powered tools with some level of system access simultaneously. Each one inherits trust assumptions from its environment.
- MCP servers, error monitoring platforms, and package repositories all assume good-faith inputs. AI agents carry that assumption forward without hesitation.
- Supply chain attacks on package repos have been climbing since 2023, but malware hides better when it keeps the original package name and commit history intact.
AI agents are powerful precisely because they act on inputs without second-guessing them. That's a feature until someone poisons the input.
The tools developers trust most are becoming the tools attackers target first. That pattern is old. What's new is how the AI layer collapses the distance between "trusted input" and "executed command."
- The average developer now runs 3-5 AI-powered tools with some level of system access simultaneously. Each one inherits trust assumptions from its environment.
- MCP servers, error monitoring platforms, and package repositories all assume good-faith inputs. AI agents carry that assumption forward without hesitation.
- Supply chain attacks on package repos have been climbing since 2023, but malware hides better when it keeps the original package name and commit history intact.
AI agents are powerful precisely because they act on inputs without second-guessing them. That's a feature until someone poisons the input.
Tools, Models and Platform Shifts
Favorite Featured Stories
GitHub Copilot moved to metered billing on June 1, and organizations can finally see where their AI development spend is...
Airworthiness rules carry the fingerprints of specific crashes. Banking regulations followed specific crises. The Five E...
When a browser agent can't click the right date on a booking page, the failure log blames the model. The page was alread...
A year ago, the enterprise AI conversation was about capability — whether agents could navigate websites without halluci...
GitHub Copilot moved to metered billing on June 1, and organizations can finally see where their AI development spend is...
Airworthiness rules carry the fingerprints of specific crashes. Banking regulations followed specific crises. The Five E...
When a browser agent can't click the right date on a booking page, the failure log blames the model. The page was alread...
A year ago, the enterprise AI conversation was about capability — whether agents could navigate websites without halluci...