Saturday, May 23 May 23, 2026

Saturday, May 23, 2026

Good morning. It’s Saturday, May 23rd, and here’s your AI morning briefing.

Today’s theme is simple: AI capability is moving faster than trust. In the last day, Anthropic shared early evidence that AI-driven cybersecurity is scaling in production. Google had to confront a very public AI search bug. Google also showed that its AI glasses are getting closer to a usable product. And on the market side, fresh reporting suggests some AI startups are still using creative math to tell a bigger growth story than the numbers may justify.

First, Anthropic. In an official update on Project Glasswing published May 22nd, the company said its roughly 50 partners have used Claude Mythos Preview to find more than ten thousand high- or critical-severity vulnerabilities across important software systems. Anthropic says some partners are seeing bug-finding rates improve by more than ten times, and it highlighted Cloudflare finding 2,000 bugs, including 400 rated high or critical, across critical-path systems. The important signal is that the bottleneck is moving. Anthropic says the constraint is no longer just discovering software flaws. It is verifying, disclosing, and patching them fast enough. That suggests the next security race will be won in workflow and remediation, not just model performance.

Anthropic also put numbers behind the open-source side of that effort. It says Mythos Preview has scanned more than 1,000 open-source projects and estimated 6,202 high- or critical-severity vulnerabilities. Of 1,752 high- or critical-rated findings already assessed, Anthropic says 90.6 percent were true positives and 62.4 percent were confirmed as high or critical. Even if those figures evolve, the takeaway is clear: frontier models are starting to change the throughput of core security work.

Second, Google’s AI Overviews had a rough day. Reporting from both The Verge and TechCrunch showed that if users searched words like disregard, ignore, or skip, Google’s AI layer sometimes answered as if the query were a chatbot instruction instead of a search term. The Verge documented responses like “Got it” and “Message received,” and later reported that Google removed the AI Overview for “disregard” while other broken responses were still visible. Google said it is aware that AI Overviews are misinterpreting some action-related queries and is rolling out a fix. It sounds like a small bug, but strategically it matters because it shows how brittle AI search still is when prompt-following and retrieval get mixed together.

Third, Google’s AI glasses look closer to a real product. In a hands-on from Google I/O, TechCrunch reported that Google demoed prototype Android XR glasses with an in-lens display, live translation, directions, contextual widgets, photo capture, and Gemini voice activation. The display-equipped version is being developed with Warby Parker, Gentle Monster, and Samsung, and Google says it will pair with both iPhones and Android phones. The key point is not that the product is finished. It clearly is not. The key point is that Google is trying to move Gemini off the phone and into ambient hardware, which means the next AI platform battle is increasingly about interface.

Fourth, the AI funding story keeps getting messier. TechCrunch reported that founders, investors, and startup finance professionals say some AI companies are blurring the line between ARR and contracted ARR, or CARR, when they talk publicly about growth. The report says investors often know the numbers are being stretched, and one VC told TechCrunch that CARR can run 70 percent higher than actual ARR. In plain English, some of the hottest AI growth stories may be counting revenue that is signed but not yet live, and may never fully materialize. As capital stays competitive, AI diligence now has to include accounting hygiene alongside model quality and distribution.

And finally, the ethics story. TechCrunch reported that the National Transportation Safety Board temporarily removed public access to its docket system after discovering that people had used AI to approximate the voices of pilots killed in a UPS crash. According to the report, users combined a spectrogram image from the accident docket with a public transcript to reconstruct cockpit audio approximations. The NTSB restored access on Friday but kept 42 investigations closed pending review. That is a reminder that as generative tools improve, even partial public records can become raw material for synthetic media.

If you zoom out, the pattern is sharp this morning. AI is getting stronger in cybersecurity and hardware, but trust is still the gating function. Search can misread simple inputs. Startup metrics can still be gamed. Public records can become reconstruction material. The winners from here are going to be the companies that combine capability with verification, workflow discipline, and institutional trust.

One business idea to watch: an AI vulnerability triage and patch-operations service for enterprises and critical infrastructure teams. It would sit between frontier bug-finding models and overwhelmed security teams, ranking findings, validating likely true positives, drafting disclosures, mapping patches to owners, and tracking remediation status. The buyers are large enterprises, cloud providers, banks, and public-sector operators. Why now is simple: Glasswing suggests discovery is scaling faster than remediation. What makes it defensible is deep integration into ticketing, SDLC, vuln-management, and audit systems, plus proprietary feedback data on which AI-found issues actually become exploitable, patchable, or false alarms.

Sources referenced in this briefing: - Anthropic, “Project Glasswing: An initial update,” May 22, 2026. - The Verge, “Google’s AI search is so broken it can ‘disregard’ what you’re looking for,” May 22, 2026. - TechCrunch, “We tried Google’s AI glasses and they’re almost there,” May 22, 2026. - TechCrunch, “How VCs and founders use inflated ‘ARR’ to crown AI startups,” May 22, 2026. - TechCrunch, “AI is being used to resurrect the voices of dead pilots,” May 22, 2026.