AI News Summary — Week 24.2, 2026

This week in AI: Anthropic pushes the frontier with its most capable model yet, Google DeepMind bets $10 million on multi-agent safety research, and NIST rolls out a mathematical framework for AI security. Here's what's new.

Anthropic Launches Claude Fable 5 — Its Most Capable Model Ever

Anthropic made waves this week with the launch of Claude Fable 5, the company's most capable model to date. The release includes both Fable 5 (the safe, public version) and Claude Mythos 5 (the unrestricted version for verified cyberdefenders).

Breaking the Benchmarks

Fable 5 sets new state-of-the-art results across nearly all tested AI capability benchmarks:

• Software Engineering: Stripe tested Fable 5 on a 50-million-line Ruby codebase and reported it completed a codebase-wide migration in a single day — work that would have taken a human team over two months.
• Knowledge Work: On Hebbia's Finance Benchmark for senior-level reasoning, Fable 5 scored highest of any model, with substantial gains in document-based reasoning, chart interpretation, and problem-solving.
• Vision: Fable 5 is the new state-of-the-art for vision tasks. It can rebuild a web app's source code from screenshots alone and even beat Pokémon FireRed with only a vision-only harness — no maps or navigation aids.
• Autonomous Research: Mythos 5 conducted a week of largely autonomous genomics research, assembling single-cell data for millions of cells across 138 animal species and designing a custom ML model that outperformed a recent Science publication — despite being 100x smaller.

The Safety Trade-Off

Releasing a model this capable comes with risks. Without safeguards, Fable 5's cybersecurity capabilities could be misused. Anthropic has implemented guardrails that route certain sensitive queries to Claude Opus 4.8 instead — these triggers affect less than 5% of sessions.

For a small group of cyberdefenders and infrastructure providers, Mythos 5 is available through Project Glasswing, a collaboration with the US government. This version has the strongest cybersecurity capabilities of any model in the world.

Pricing

Fable 5 and Mythos 5 are priced at $10 per million input tokens** and **$50 per million output tokens — less than half the price of Claude Mythos Preview.

Google DeepMind Bets $10M on Multi-Agent Safety

Google DeepMind, together with Schmidt Sciences, ARIA (the UK government's moonshot agency), the Cooperative AI Foundation, and Google.org, announced a $10 million funding pot for researchers to study the behavior of multi-agent AI systems.

Why It Matters

The concern: as millions of AI agents get deployed across the economy and begin interacting with each other, we could hit a tipping point where imagined risks become real.

"We've got this digital commons that is integral to how society works, and you really want to ensure that this doesn't descend into absolute anarchy," said James Fox of Schmidt Sciences.

The risks include:

• Supercharged scams: AI agents that can collaborate to run sophisticated fraud operations
• Prompt injection as malware: AI agents fed malicious instructions, turning them into self-guiding malware
• Cyberattacks at scale: Mass automated attacks powered by coordinated agent swarms

The Research Gap

"The main issue is that there just isn't really a field of research for multi-agent safety yet," said Google DeepMind's Rahul Shah. "And we would like there to be."

The funding aims to kick-start academic research outside tech companies — work that can look far into the future without the pressure of product timelines.

NIST Provides Mathematical Proof for Continuous AI Security

The National Institute of Standards and Technology (NIST) published a mathematical proof supporting a transition to continuous-monitor-and-update security models for AI systems.

The proof establishes a formal mathematical framework for why AI systems require ongoing monitoring rather than one-time security assessments. Traditional cybersecurity assumes fixed software doing fixed things — AI agents reason, improvise, and can be hijacked by a single hidden sentence in a document they process.

This validates the "zero trust for AI agents" approach that Anthropic published guidance on last month.

What's Hot This Week

• Claude Fable 5: Anthropic's most capable model ever, with new safety guardrails
• Multi-agent safety: Google DeepMind commits $10M to research
• NIST security proof: Mathematical framework for continuous AI monitoring
• Mythos 5: Now deployed through Project Glasswing for US cyberdefense

That's a wrap for Week 24.2. Stay tuned!

Sources: Anthropic, MIT Technology Review, NIST, TechCrunch