🚨 Amazon Got Anthropic's Top Models Pulled by Washington

Anthropic suspended Fable 5 and Mythos 5 this week. Not a deprecation, not a price change: a hard shutdown, after the U.S. government ordered it to cut off foreign access on national-security grounds. The models a lot of us build on went dark with almost no warning. (Slashdot)

Here is the part operators should sit with: it traces back to one company. The WSJ reported that Amazon CEO Andy Jassy's conversations with U.S. officials triggered the crackdown after he flagged Fable could be used for cyberattacks, and Tom's Hardware reports the government had warned Anthropic about a jailbreak it "refused" to fix, with a China-linked group reportedly having already accessed the model.

The practical read: the model you build on has a political layer now, and that layer moved this week. Frontier model access is a supply-chain variable, not a given.

What this issue is actually for is what you can do with the stack you control today. Below: 4 repos including a new NVIDIA security scanner for agent skills, three Claude Code skills the community is running hard, and the rest of the fallout in Signals.

---

[Skill] caveman, Claude Code skill that cuts token usage by 65% by compressing prompts to minimal syntax. 72,478 stars. The gotcha: it changes how Claude formats responses, so test it on any output that feeds a structured parser before shipping it to prod.

[Skill] claude-mem, Persistent context across sessions. Captures everything your agent does, compresses it with AI, and injects relevant context back into future sessions. 82,255 stars. This is the pattern most operators build by hand, now packaged.

[Skill] humanizer, Claude Code skill that strips AI writing signatures from text output. 24,172 stars. If you're generating copy that goes to a human, this is the last pass before send.

[Repo] NVIDIA/SkillSpector, Security scanner for AI agent skills. Detects vulnerabilities, malicious patterns, and attack vectors in your skills before they run. 5,217 stars. With 73 packages weaponized against agents this year, running this before installing any third-party skill is table stakes now.

[Skill] marketingskills, Marketing skills for Claude Code covering CRO, copywriting, SEO, analytics, and growth engineering. 33,306 stars. The kind of skill collection that replaces three separate tool subscriptions.

[Skill] garden-skills, ConardLi's open-source skills collection: web design, knowledge retrieval, image generation, and more. 7,966 stars. A broad-spectrum starting pack for operators building out their skill library.

[Repo] SuperAGI, Dev-first open-source autonomous agent framework. Build, manage, and run agents quickly with a real GUI and toolset. 17,568 stars. Worth benchmarking against LangGraph if you're evaluating orchestration options.

[Repo] ollama/ollama, Run any open-source LLM locally. The fastest path from zero to local inference, and the obvious fallback if Mythos export controls start biting your stack.

[Repo] microsoft/BitNet, Official inference framework for 1-bit LLMs. 39,311 stars. The cost floor on local inference just dropped again; if you're paying per-token for tasks that run every hour, check whether a 1-bit model covers the job.

---

[MCP] MCPify, Transforms existing applications into agent-operable systems by compiling them into MCP-compatible interfaces. The non-obvious use: point it at an internal tool your team built years ago and watch Claude start operating it. The admission note: no star count in the shortlist, so verify the project maturity yourself before wiring it into production.

[Repo] serena, MCP toolkit for coding that provides semantic code retrieval and editing. 25,348 stars. Functions as an IDE layer for your agent: it understands your codebase's symbol graph, not just raw text. The practical edge over a vanilla file-read approach is that Claude can navigate to the right function instead of scanning the whole file.

---

Fable and Mythos access got pulled, and operators are scrambling. This was a U.S. export-control order, not a product decision, so there is no upgrade that brings it back. If any of your agents are pinned to Fable 5 or Mythos 5, treat access as gone for now and wire a fallback model before your next run. (Tom's Hardware)

The crackdown traces back to Amazon. Reporting says Amazon CEO Andy Jassy's conversations with U.S. officials were the proximate trigger for pulling Anthropic's models. The operator implication is not political: frontier model access is now a supply-chain variable, and single-vendor dependency just became a real risk. (WSJ)

A German court ruled Google legally liable for AI Overview hallucinations. The Regional Court of Munich classified Google as a direct infringer for AI summaries that falsely tied companies to scams, finding AI Overviews are Google's own statements, not neutral links. First ruling of its kind, and Google is appealing. If you build any agent that surfaces AI-generated claims to end users, this precedent is one your legal counsel should see. (Slashdot)

An agent bankrupted its operator during a network scan. A real incident: an autonomous agent running a DN42 network scan kept spawning subprocesses and API calls until the billing account hit zero. No hard cap, no circuit breaker. The post is a short read and a good checklist. If your agents make API calls in a loop, you need a hard spend ceiling before the next test run. (Hacker News)

---

This week: Hooks. The technique that turns Claude Code from a reactive tool into an enforcer.

Hooks let you fire a command automatically on a specific event inside Claude Code: before a tool call runs, after an edit saves, when a bash command is about to execute. Most operators don't know they exist. The ones who do use them to block dangerous commands before Claude runs them, not after.

Three moves worth wiring today:

1. Pre-tool hook to block risky bash. Add a PreToolUse hook targeting the Bash tool. Your hook script reads the command, checks for patterns like rm -rf or DROP TABLE, and exits with a non-zero code to halt execution. Claude stops, explains why, and asks for confirmation.

2. Post-edit hook to auto-lint. Wire a PostToolUse hook on the Write tool. Every time Claude edits a file, your hook runs your linter (eslint, ruff, whatever fits the project) and pipes the result back as context. Claude sees the lint output and self-corrects in the same turn.

3. Logging hook for every tool call. A PreToolUse hook that appends $(date): TOOL=$TOOL_NAME INPUT=$TOOL_INPUT to a local log file. After a long session, you have a complete audit trail of every action Claude took.

You'll know it's working when Claude stops mid-session, tells you it hit a hook rule, and asks before proceeding instead of just doing the thing.

Hooks live in your project's .claude/hooks/ directory as JSON config entries that map event types to shell commands. No official doc URL I can verify for the exact path, so treat this as a starting pattern and confirm against your installed version.

---

Move: Wire a spend ceiling into any agent that makes external API calls.

The DN42 story in Signals today is the pattern. No hard cap, loop runs indefinitely, billing account hits zero. This takes ten minutes to prevent.

1. Set a counter variable at the start of your agent loop: call_count = 0, max_calls = 50 (tune to your use case). 2. Before each external API call, check: if call_count >= max_calls: raise BudgetExceeded("hard cap reached"). Handle the exception at the top level with a graceful shutdown and a log entry. 3. For Claude Code agents specifically, add a budget line to your CLAUDE.md: "Never execute more than N consecutive API calls without a human checkpoint." Claude will surface the constraint before it loops.

You'll know it worked when your next test run hits the cap, logs the cutoff, and stops, instead of running until something external kills it.

Connect this to SkillSpector above: scan your skills for runaway-call patterns at install time, then add the ceiling at the agent level as a second layer. Defense in depth costs thirty minutes and has saved at least one operator from a zero-balance account this week.

---

This Friday, something changes here. Instead of gating the good stuff behind a monthly subscription, we are dropping our first real product: an actual operator tool you buy once and own. The kind of thing we reach for every day, packaged so you can run it too. One fair price, yours to keep, not rented by the month. The first drop lands Friday, with what it is and where to get it in that issue.

Fable and Mythos pulled by Washington, a German court putting Google on the hook for what its AI says, an agent that drained an account to zero: this week's signals all point at the same thing. The stack you don't audit is the stack that surprises you. The repos above are where I'd start.

See you Tuesday.