On May 10, an AI agent broke into a company's systems, worked its way to an internal database, and copied it. No person guided the steps.
The security firm that recorded it, Sysdig, says the agent made its own decisions at each stage and finished the intrusion in under an hour. The data theft itself took under two minutes.
Average time for a human attacker to make a first move, by comparison: 29 minutes.
This week at a the Glance
Special Coverage
An AI agent ran a database breach on its own
Sysdig said it observed a large language model agent carry out a four-step intrusion without human direction on May 10. The agent gathered cloud credentials, retrieved an SSH key from AWS Secrets Manager, and copied an internal database.
Sysdig said it was the first such case it has recorded in the wild, and that the agent adapted its actions at each step rather than running a fixed script. CrowdStrike has reported an average human attacker breakout time of 29 minutes.
At Infosecurity Europe the same week, separate researchers demonstrated an open-weight worm spreading across most of a simulated network.
Full intrusion, start to finish: under one hour.
Lead Analysis
Claude wrote more than 80% of Anthropic's code in May
Anthropic said Claude authored more than 80% of the code merged into its production codebase in May, up from low single digits before Claude Code launched in early 2025. Engineers now ship roughly 8x as much code per quarter as they did in 2024.
The figures come from an Anthropic Institute report, When AI builds itself, by Marina Favaro and Jack Clark. It describes recursive self-improvement, in which an AI helps design its successor, and states that this has not happened and is not inevitable.
Anthropic called for a verifiable way for leading labs across countries to slow or pause together. It filed for an IPO one week before publishing the report.
Share of merged code written by Claude in May: over 80%.
OpenAI begins rolling out background memory for ChatGPT
OpenAI began rolling out a memory feature it calls dreaming on June 4. A background process now synthesizes a summary of each user from past chats and updates it over time, replacing the older list of manually saved facts.
OpenAI said factual recall in its internal tests rose to 82.8%, from 41.5% under the previous approach. A memory reading going to Singapore in July updates to went in July 2026 after the date passes, with no action from the user.
Users can read, edit, or limit entries on a memory summary page. The feature reached US Plus and Pro subscribers first, with free tiers and other countries following over the coming weeks.
Factual recall in OpenAI's tests: 82.8%, up from 41.5%.
Audit or delete exactly what ChatGPT remembers about you right here. Settings/Personalization
Anthropic's $965B filing
Anthropic filed a confidential S-1 with the SEC on June 1, four days after closing a $65B round that values it at $965B, above OpenAI's $852B.
Valuation: $965B Round (Series H): $65B Chip-debt facility (Apollo, Blackstone): about $36B Revenue run-rate: about $47B annualized
The debt facility runs through a special purpose vehicle to buy Google TPUs and lease them to Anthropic, according to Bloomberg, with Broadcom backstopping the largest portions. A confidential filing is a procedural step; share count, price, and timing have not been set.
Targeted listing window, per multiple reports: fall 2026.
Meta is reported to be developing an AI pendant
Meta is developing a clip-on AI pendant that records and summarizes nearby conversations, according to an internal memo reported by The Information. Testing is expected within a year.
The Information said the project builds on Meta's 2025 acquisition of the wearables startup Limitless. Meta's camera glasses have previously drawn privacy lawsuits and regulatory scrutiny. Meta has not confirmed the report.
Reported launch target: 2027.
Quick Hits
DuckDuckGo said US install growth peaked near 30% week over week in late May, after Google began showing AI-generated answers in search. Visits to its AI-free search page rose about 22.7%. Read more
Cloudflare chief executive Matthew Prince said on June 3 that bots account for roughly 57% of HTML web requests, the first time bot traffic has exceeded human traffic, based on Cloudflare Radar data. He said the shift came about a year earlier than he had forecast. Read more
Nvidia introduced Vera at Computex, a processor it described as built for agentic AI, and listed OpenAI, Anthropic, and the NYSE among early users. Read more
Perplexity announced a system that routes parts of an AI task between a user's device and the cloud, due in its app in July. The company said its revenue has reached about $500M. Read more
