🎯 Top 3 Things to Know
1. Google's Threat Intelligence Group disrupted the first documented case of an AI model finding and weaponizing a zero-day vulnerability in an active attack. A criminal group used an unaffiliated model called OpenClaw to discover and exploit an unknown flaw in a widely deployed open-source sysadmin tool, building a Python script that bypassed two-factor authentication. Google says it caught the operation before it became a mass exploitation event, then worked with the vendor on a coordinated disclosure. The friction this exposes is real. Vulnerability discovery has long been a labor and skill bottleneck. AI has now compressed both. Defenders will need to assume that adversaries are scanning their stacks with frontier-grade tooling, and that pre-disclosure windows will shrink. Worth watching how patch SLAs and bug bounty pricing adjust over the next quarter. Bloomberg · Google Cloud
2. OpenAI launched the OpenAI Deployment Company, a separately capitalized $4 billion enterprise services arm with TPG, Bain Capital, Advent, and Brookfield as co-leads, and acquired Tomoro to seed the team. The new company embeds Forward Deployed Engineers inside enterprises to redesign workflows around frontier models, then handle the integration work that has historically slowed pilot-to-production transitions. The Tomoro acquisition adds about 150 deployment specialists with existing accounts at Mattel, Red Bull, Tesco, and Virgin Atlantic. The structure echoes how Palantir built its services moat. It also mirrors Anthropic's growing roster of integrator partnerships. The signal for the broader market is that frontier labs increasingly view deployment friction, not model capability, as the binding constraint on enterprise revenue. Worth watching whether margins on deployed AI compress as labs and integrators converge on the same accounts. OpenAI · PYMNTS
3. EU lawmakers reached an AI Omnibus agreement that softens and stretches several AI Act timelines, the first material rollback since the law took effect. Council and Parliament negotiators settled at 4:30 a.m. on May 7. The deadline for national AI regulatory sandboxes slips to August 2027. The grace period for transparency on AI-generated content shrinks from six months to three, with a December 2 deadline. Enforcement of certain general-purpose AI obligations centralizes further at the EU AI Office, and small mid-cap firms get the same compliance breaks as SMEs. The Commission opened a public consultation on transparency guidelines a day later. The shift is pragmatic. Industry has been arguing for two years that some original deadlines collided with the pace of model releases. Worth watching whether the U.S. state laws coming online this summer (Colorado in June, others queued) follow the same pattern of post-enactment softening. Council of the EU · Tech Policy Press
🚀 Frontier Models & Features
Quiet day on lab releases. The week's main event sits ahead: Google I/O opens with the Android Show I/O Edition today and the main keynote on May 19, where Gemini 4 is widely expected alongside Project Astra and Veo updates. GPT-5.5 Instant remains the ChatGPT default following last week's rollout. Google I/O · BusinessToday preview
🔬 Research Worth Reading
STALE: Can LLM Agents Know When Their Memories Are No Longer Valid? (Chao, Bai et al.). arXiv 2605.06527
- TL;DR: A new benchmark stress-tests whether an agent notices when a stored fact has been silently invalidated by later context. Three probes: detect that an old belief is stale, refuse a question that presupposes the stale state, and proactively act on the updated state.
- Stat: Best frontier model evaluated reaches only 55.2% accuracy across the 1,200-query benchmark. Most failures come from "implicit conflict," where invalidation requires inference rather than explicit negation.
- Apply it: Add a small set of implicit-conflict scenarios to your agent eval suite. If your retrieval pulls a memory and the model never asks whether it is still true, you have a bug class your current tests are not catching.
Learning Agent Routing From Early Experience (Wang, Qiu, Qi et al.). arXiv 2605.07180
- TL;DR: Training-free router that decides per query whether to answer with a direct LLM call or escalate to a full agent. Builds a small experience memory from running both on a seed set, then retrieves similar past cases at inference time to guide the routing decision.
- Stat: Cuts inference time 60.6% versus always-agent, while improving accuracy 28.6% over direct LLM. Beats prompt-only routing by 37.9% on average.
- Apply it: Before adding more sophistication to your agent, measure how often a single LLM call would have gotten the same answer. A retrieval-based router on a few hundred labeled examples is often the cheapest latency win available.
🏢 Enterprise in the Wild
ServiceNow and Accenture launched a Forward Deployed Engineering program to scale agentic AI on the ServiceNow AI Platform. Joint teams will design and ship workflow agents directly inside customer environments. The structure parallels OpenAI's new Deployment Company and confirms that integrator-led delivery is becoming the default enterprise motion. Accenture
🛠️ Tooling & Ecosystem
Anthropic's Claude models are now generally available in Microsoft Foundry, including Opus 4.7, Opus 4.6, and Sonnet 4.6 with 1M-token context. Foundry traffic joins Bedrock and Vertex AI as a third hyperscaler routing path for Claude. Tenant controls and EU data residency remain on Microsoft's published roadmap. Microsoft Azure
⚖️ Policy & Regulation
The U.S. Center for AI Standards and Innovation (CAISI) signed pre-deployment evaluation agreements with Google DeepMind, Microsoft, and xAI, extending the framework already in place with OpenAI and Anthropic. CAISI will run capability and security assessments on frontier models before public release. The arrangements are voluntary and bilateral, not regulatory, but bring the major U.S. labs under a common pre-release testing regime for the first time. Reuters via DigiTimes
📌 Watch List
- AI-assisted vulnerability discovery moving from theory to incidents in the wild, and how patch SLAs respond.
- Forward-deployed engineering as the new enterprise sales motion, now adopted by OpenAI, ServiceNow, and Accenture.
- Cost-aware agent routing as a cheap latency win before sophistication is added to a stack.
- Agent memory validity. STALE adds quantitative weight to a problem most production agents currently ignore.
- EU AI Act timelines softening, and whether U.S. state laws follow the same pattern.
- Google I/O on May 19. Gemini 4 and Astra updates the most likely shape-of-the-week items.