US
Government Suspends Anthropic’s Fable 5 and Mythos 5
A US export control directive ordered Anthropic to suspend global
access to Fable 5 and Mythos 5 — the company’s newest models — after the
government discovered what it described as a method to bypass safety
features. The trigger: a “narrow, non-universal jailbreak” involving
code vulnerability analysis. Anthropic disputes the severity, arguing
such capabilities are widely available in other models and that applying
this standard would “essentially halt all new model deployments.”
Whatever the technical merits, this is a first-of-its-kind regulatory
precedent for frontier AI — and it lands the same week an outside team
made the export-control argument concrete by replicating one of the
relevant capabilities for $1,000.
21
FFmpeg Zero-Days for $1,000 — Roughly a Tenth of Anthropic’s Bill
Security firm depthfirst’s autonomous agent discovered 21 zero-days
in FFmpeg for around $1,000 — roughly 10% of what Anthropic reportedly
spent using Mythos for similar work. The bugs include heap and integer
overflows aged 3 to 23 years. The most critical hits the AV1 RTP
depacketizer, where a single 183-byte packet can redirect execution —
exposing any system that accepts RTSP streams, including media ingest
pipelines and surveillance infrastructure.
“Shepherd’s
Dog”: A Complete Browser Game from a Single Fable 5 Session
Before the suspension landed, Claude Fable 5 produced “Shepherd’s
Dog” — a complete browser game in one 45-minute session: 2,319 lines of
HTML/JavaScript, zero external dependencies, about €20 in tokens, and,
per the developer, “exactly how I imagined it.” The interesting bit
isn’t the artefact, it’s the workflow disappearing: no task
decomposition, no iterative refinement loop, no scaffolding.
A
Practical Local Coding Agent on an M1 Max — 72 Tokens/s
A comprehensive walkthrough shows how to run a usable local coding
agent on macOS with llama.cpp, Metal, Gemma 4 26B-A4B, and speculative
decoding via an MTP draft model — hitting 72.2 tokens/s, a 24% lift over
the base model. The guide covers multimodal screenshot ingestion and
benchmarks llama.cpp outperforming MLX on M1 Max. Local coding
assistance has tipped from “demo” into “default.”
WASI 0.3
Brings Native Async to WebAssembly Components
WASI 0.3 integrates async natively into WebAssembly Components: the
host now controls a unified event loop across components, complex
three-step patterns are replaced with direct async func
exports/imports, and both stackless and stackful coroutines are
supported across languages. Component chaining could collapse
inter-service call latency from milliseconds to nanoseconds.
Nix Flakes
vs Guix — One Big Lever vs Many Small Ones
A detailed side-by-side argues the philosophical split clearly:
Flakes are “one big feature that solves a bunch of problems at once” —
per-project inputs, automatic flake.lock,
standardized entry points, built-in discoverability. Guix is “a
collection of smaller, orthogonal tools” — system-wide channels, manual
pinning, with guix time-machine for historical
reproducibility and grafting for fast security updates. Useful as a
clarifying read for anyone who’s drifted between the two.
Why CPU Speed Is a Geometry
Problem
A draft chapter from an upcoming book on efficient C++ programming
lays out the physical story behind cache latencies: register ~1 cycle,
L1 ~3, L2 ~10–15, L3 ~30–70, RAM ~200–300. The gap isn’t a light-speed
limit — it’s parasitic capacitance — and it’s why vectors keep beating
linked lists in practice regardless of what the asymptotic analysis
says.
APLR(1): A Simpler
Path to Compact LR(1) Parsers
The APLR(1) (Adequacy Preservation LR(1)) algorithm offers a
conceptually simpler alternative to IELR(1) for generating compact LR(1)
parsers. IELR(1) starts from LALR(1) and eliminates inadequacies through
state splitting; APLR(1) starts from canonical LR(1) and preserves
adequacy during state remerging. The Hocc implementation is 700 lines vs
IELR+(1)’s 1,600, and the failure modes are “messy and loud” rather than
“clean and silent” — a real win for parser-generator maintainers.
SpaceX employees celebrate during a bell ringing ceremony for the IPO of
SpaceX at the Nasdaq MarketSite, June 12, 2026. Photo: Deseret News
Fire at the Nizhnekamsk oil refinery
in Tatarstan, June 12, after a Ukrainian drone strike on the TANECO
facility. Photo: Militarnyi