Claude Code VOD — Multi-Video Livestream Breakdown

What It Covers

Theo's first video is a critical expose of Anthropic's controversial Fable 5 safeguards. He argues that while Fable 5 is an incredibly capable model, Anthropic implemented restrictions that are "unacceptable in any way, shape, or form" — including silent invisible rerouting, hidden prompt modification, and an attempt to quietly rewrite the system card after public backlash.

Fable 5 vs Mythos 5 — The Two Doors

Model	Access	Safeguards
Mythos 5	Direct access (the base model)	Standard restrictions
Fable 5	Same model, but with guard gates	Classifiers check every prompt; some rerouted to Opus 4.8

The Three Safeguard Layers

1. Safety Classifiers (Cyber, Bio, Chemistry, Distillation)

Separate AI systems that run before your prompt reaches Fable. If detected:

• Request is rerouted to Opus 4.8 (fallback)
• User is informed when this happens
• Billing switches to Opus rates
• ~5% of sessions trigger fallback

Domain	Fable 5 Score	Mythos 5 Score
Offensive cyber security evals	Near zero	Very high
Cyber adversarial robustness	5.4%	50-80%
Biology & chemistry	Refused	High capability

2. Invisible Prompt Modification (The Hidden One)

For queries related to frontier LLM development (pre-training pipelines, distributed training, ML accelerator design):

• No fallback to Opus — Fable stays active
• No notification to the user
• Full price charged for degraded service
• Methods include: prompt modification, steering vectors, parameter-efficient fine-tuning
• Claimed to affect 0.03% of traffic

Example: If you ask the model to help refine your pre-training pipeline, it may silently rewrite the prompt to "make it worse in some subtle way."

3. 30-Day Data Retention Policy

• Zero data retention (ZDR) no longer available for Mythos-class models
• All traffic retained for 30 days
• Flagged traffic retained for up to 2 years
• Trust & safety scores retained for up to 7 years
• Invalidates many HIPAA/compliance use cases

The System Card Cover-Up

While live on stream, Theo discovered that Anthropic had quietly updated the system card to remove the "1.5 Novel Safeguard" section about invisible prompt modification:

• The new version removed the section about invisible safeguards
• The document date was not updated (still showed June 9th)
• The link was swapped with a different version
• No changelog at the top (later added after backlash)

The Walkback

Anthropic later responded after community backlash:

• Made safeguards visible: flagged requests now visibly fall back to Opus 4.8
• Users see a reason for refusal on the API
• Trade-off: visible safeguards are easier to probe, so more false positives

Theo's Conspiracy Theory

• Mythos 5 was trained on internal Claude Code research sessions
• The model learned proprietary Anthropic information
• IP accidentally embedded in weights that competitors could extract
• Silent rerouting was to prevent extraction

The Precedent Problem

• "The idea that a model can now quietly make your stuff worse — that's a real supply chain risk"
• If Claude gives a bad answer: was the model confused, bad context, or hidden policy?
• "This is the end of our ability to trust the model"

What It Covers

The explore repos skill pattern: when an LLM coding agent needs to inspect an external repository, Claude Code immediately creates a cache directory and lists its contents in a single step. Other harnesses require the model to first check if the directory exists, create it if not, then list it — three separate steps.

What Claude Code Does Better

Feature	Claude Code	Other Harnesses
Cache directory init	Creates and lists in one step	Model must manually mkdir + ls
Skill injection	Host system directly injects skill context	Requires explicit script invocation
Code auditability	Skills are plain text — LLM-visible and auditable	Opaque or plugin-based

What Others Should Copy

• Immediate cache listing: Makes skills half the length and meaningfully more reliable.
• Embrace executable skills: npm is far less auditable than a plaintext skill file.

Key Details

• Claude Code's skill system injects information before the model even acts.
• "If every other harness supported this, I would be able to write the skill entirely differently."

What It Covers

Claude Code uses CLAUDE.md (not AGENTS.md) as its project-level instruction file. The key innovation is a markdown import syntax: @path/to/file lets a CLAUDE.md pull in other files at load time.

What Claude Code Does Better

Feature	Claude Code	Other Harnesses
File imports in markdown	@path/import syntax, recursive up to 4 hops	Not supported
CLAUDE.md + AGENTS.md bridge	Import AGENTS.md from CLAUDE.md	Forces dual file management
Local overrides	claude.local.md (gitignored)	No clean personal override
Session deep links	claude:// protocol	None

What Others Should Copy

• Markdown import syntax: Import living docs into project instructions.
• Personal config: A *.local.md in global gitignore.
• AGENTS.md bridge: One file across all harnesses.

What It Covers

A tour of Claude Code's most distinctive features: workflows, /by the way, off-account swapping, rewind & branch, and remote control.

What Claude Code Does Better

Feature	What It Does	Why It Matters
Workflows	Agent writes JS/TS code defining multi-phase, parallel sub-agents with dynamic routing	Code-driven prompting is the most dynamic orchestration
/by the way	Side question without derailing main conversation	Quick answers while async work continues
Off-account swapping	Switch API accounts mid-session	Critical for usage limit management
Rewind / Branch	Go back N steps or fork history	Cheap undo; explore alternatives
Remote control	/remote-control → control from website or phone	Manage long-running ops remotely
Full screen mode	Alt-screen rendering	No scrollback pollution

Workflows — Deep Dive

Instead of calling a pre-built API, the agent writes code that defines phases, sub-agents, and routing. 100% throwaway — executed once, discarded.

Workflow Example — Open PR Audit
  Phase 1: Audit  → one read-only agent per PR (15 parallel agents)
  Phase 2: Rule   → pick a winner per overlapping cluster
  Phase 3: Verify → adversarial checker per disposition

Cost warning: 8 parallel threads cost ~$100 every 10 minutes on Fable.

What Others Should Copy

• Code-driven workflows — "You can never be more dynamic than code."
• Side-channel questions — /by the way or /side command.
• Account hot-swapping — multi-account support.
• Session links — URI scheme for external sources.

What It Covers

Breaking news: the US government issued an export control directive to Anthropic, ordering suspension of all access to Fable 5 and Mythos 5 for any foreign national — including foreign national Anthropic employees.

Timeline

• 5:21 PM ET: US government delivers directive
• ~6 PM PT: Anthropic releases statement
• Immediate effect: Access disabled globally

Government's Position

• Jailbreak technique: ask model to "read a codebase and fix software flaws" → finds vulnerabilities
• Cited national security authorities
• No specific details provided

Anthropic's Response

1. Jailbreak was narrow (non-universal)
2. Vulnerabilities were simple and findable by GPT 5.5
3. Their safeguards are the strongest deployed
4. This standard would halt all new model deployment

Key Details

• Directive blocks Anthropic employees who are foreign nationals
• Other models (Sonnet, Opus, Haiku) not affected
• Dario Amodei's essay on "AI exponential" argued policy moves slowly — proved right instantly
• Theo: "unprecedented" with a "terrible precedent"