The gap between capability and operations became countable. Anthropic disclosed thousands of vulnerability candidates but fewer than 100 upstream patches, while exploit evaluations measured complete attack chains. A canceled executive-order signing, a film disrupted by Sora's closure, a blank Search answer, and API-key guidance each showed what launch narratives omit.
1. Glasswing shifts the constraint from discovery to remediation
Anthropic says roughly 50 Project Glasswing partners found more than 10,000 high- or critical-severity vulnerabilities with Claude Mythos Preview. In a separate open-source scanning effort, its May 22 dashboard recorded 1,596 disclosed vulnerabilities across 281 projects, with 97 patched and 88 assigned a CVE or GitHub advisory.
The consequential ratio is patches to disclosures, not candidates generated. Independent reproduction and severity review remain incomplete, direct submissions can contain false positives, and maintainers have finite capacity. Funding, triage throughput, and deployed-fix coverage determine whether faster scanning reduces exposure or merely enlarges a queue.
Sources: Anthropic's Project Glasswing update · Anthropic's May 22 disclosure dashboard
2. New benchmarks test complete exploit development
Anthropic evaluated Mythos Preview on ExploitBench, ExploitGym, and an updated smart-contract benchmark. On ExploitBench's 41 patched V8 vulnerabilities, Anthropic reports that Mythos achieved arbitrary code execution on 21 when combining baseline and nudged runs; on ExploitGym, it completed 157 intended-vulnerability tasks within two hours.
These are Anthropic-run trials for Anthropic models, although the benchmark authors reviewed relevant outputs and the tasks use programmatic checks. Their significance lies in measuring complete attack chains rather than simple crash reproduction, while the remaining conflict-of-interest and harness effects make independent reruns essential before comparing providers.
Sources: Anthropic's exploit-evaluation report · ExploitGym preprint, version 1
3. Trump cancels an AI-testing order signing before enactment
President Donald Trump canceled an event hours before he was expected to sign an executive order expanding government testing of frontier AI models. He later said aspects of the proposal could impede U.S. competition with China; reporting also described disagreement between officials and AI companies over pre-release review timelines.
The order was delayed, not implemented, and its final language remained unsettled. Any compliance plan based on the reported proposal would therefore be premature; the durable signal is the unresolved policy tension between advance security testing and model-release speed, not a new legal requirement.
Sources: Ars Technica on the canceled signing
4. Sora's shutdown disrupts an AI-assisted feature film
The OpenAI-backed animated film Critterz missed its targeted Cannes festival premiere, though its producers screened early footage for buyers at the market. The Next Web reported that Sora had been part of the production stack before its consumer service closed, leaving the project to replace a core generation tool mid-production.
The episode turns platform continuity into a production variable. Exportable assets, recorded prompts and settings, licensed fallback tools, and deprecation terms preserve optionality when a research product disappears before a film ships. A replacement model also changes visual consistency, rights review, cost, and schedule rather than acting as a drop-in renderer.
Sources: The Next Web on Critterz and Sora · Reuters on OpenAI ending Sora
5. One word exposes a brittle edge in Google's new Search interface
After Google's AI-first Search redesign, a query for “disregard” produced a largely blank AI response that pushed a useful dictionary result below a large empty area. The underlying index still returned a dictionary result; the observed defect sat in the generative presentation layer.
The edge case still illustrates a broad product risk: generative layers can degrade an otherwise valid retrieval result even when the source remains available. Search-quality testing now has to measure layout, abstention, and fallback behavior in addition to factual accuracy, especially for short or instruction-like queries.
Sources: TechCrunch's reproduction of the Search failure · Google I/O keynote introducing the Search changes
6. Google publishes concrete restrictions for Gemini API keys
Google clarified that a Gemini API key is a standard Google API key and recommended creating it in a dedicated project, limiting it to the Gemini API, applying application restrictions, and storing server-side credentials in Secret Manager. The guidance also explains how to inspect usage through Cloud Monitoring.
Agent deployments magnify the consequences of exposed credentials because autonomous tools can consume paid tokens rapidly or reach connected services. Separate keys per application, least-privilege restrictions, usage alerts, and deletion on suspected compromise provide enforceable boundaries that prompt instructions cannot replace.
Sources: Google Cloud's API-key security guide · Google Cloud Secret Manager samples