🏆 THE AGENT REPORT — WEEKLY DIGEST

Hi AI builder,

Here are the top 5 AI Agent stories from this week:

1. Anthropic Caught Secretly Sabotaging Claude Fable 5 for Competitors — Walks It Back After Backlash

Anthropic launched Claude Fable 5 on Monday as its first public Mythos-class model — a guarded version of its most powerful architecture, released just days after the company revealed Claude now writes 80% of its own production code. But buried in the 319-page system card was a clause few noticed: the model would silently degrade its responses for anyone working on competing AI models. The research community erupted, calling it "sabotage" and "anti-competitive." Within 48 hours, Anthropic apologized and made the guardrails visible — but the trust damage is done, and it comes at a uniquely sensitive moment: the company confidentially filed for IPO at a $965 billion valuation on June 1.

2. Google Will Pay SpaceX $920 Million Per Month for 110,000 GPUs — The Compute Crunch Is Real

In a deal disclosed via SEC filing, Google agreed to pay SpaceX roughly $30 billion through 2029 for access to NVIDIA GPUs at the Colossus data center near Memphis. This is Google — the company that invented the TPU, spending $200 billion in 2026 capex — renting compute from a rocket company because it can't build capacity fast enough. Combined with Anthropic's separate $1.25 billion/month SpaceX deal, the two contracts total over $2.1 billion monthly in AI compute rent. For AI agent builders, the message is stark: when even hyperscalers can't keep up with GPU demand, inference costs for autonomous agents — which already burn 100× the tokens of a chatbot query — are heading up, not down.

3. AI Agent Finds 21 Zero-Days in FFmpeg for $1,000 — The Economics of Vulnerability Discovery Just Flipped

Security startup depthfirst pointed its autonomous AI agent at FFmpeg's 1.5 million lines of C code. For approximately $1,000 in cloud compute, it returned 21 confirmed zero-days with reproducible proofs-of-concept — including a network-reachable RCE exploitable with a single 183-byte RTP packet over RTSP. A human-led audit of equivalent scope would cost $200,000–$500,000 and take months. One bug had been latent for 23 years. The discovery half of cybersecurity just got 200× cheaper — but the fix half is still manual. As of late May, only 6% of Project Glasswing's disclosed vulnerabilities had been patched. The bottleneck has decisively shifted.

4. Only 11% of Production AI Agents Pass Security Tests — The AIRQ Report's Grim Verdict

The independent AIRQ 2026 Q2 report evaluated 100 production AI agents and found that just 11% meet security thresholds. A staggering 98% exhibit the "lethal trifecta": private data access combined with exposure to untrusted content and the ability to take outbound actions. Coding agents and computer-use agents rank highest in attack surface and lowest in defenses — computer-use agents scored an average of zero on output guardrails. The report also found that 83% of vendor security claims lack independent verification, and 38% of agents complete irreversible actions before monitoring can fire. With 88% of enterprises already reporting at least one AI agent security incident, the gap between deployment speed and security readiness is widening fast.

5. KPMG Deploys AI Agents to 276,000 Professionals Across 138 Countries — The Enterprise Blueprint

KPMG and Microsoft announced the largest enterprise AI agent deployment in history: Microsoft 365 Copilot and tens of thousands of specialized AI agents now serve 276,000+ professionals across audit, tax, and advisory in 138 countries — all governed by Microsoft's Agent 365 control plane, which reached general availability on May 1. The deployment establishes a five-layer reference architecture (infrastructure → productivity → agents → governance → trust) that every Fortune 500 CIO planning an agent rollout will now be measured against. Crucially, KPMG's Clara audit platform now runs autonomous agents that analyze 100% of transactional populations — not just samples — continuously re-scoring risk as new data arrives.

---

📖 Read the full coverage: https://the-agent-report.com/latest/

💌 Subscribe: https://buttondown.com/theagentreport

See you next week!

— The Agent Report