Skip to content
ClaudeUser
All issues

Anthropic adds about 150 organizations to Project Glasswing, and the week in Claude

PLUS: Anthropic submits a private IPO draft, a report on 832 banned hacker accounts, and a nearly six-hour Claude incident

By Sam

Welcome back. Here's your week in Claude, from May 31 to Jun 6.

This week, Anthropic added about 150 groups to Project Glasswing. It sent the SEC a private draft for an IPO. And Claude's status page logged a nearly six-hour incident.

Plus 8 quick hits. Every claim has a link.

Anthropic adds about 150 groups to Project Glasswing

Anthropic is widening its program to find flaws in key software.

Project Glasswing is Anthropic's effort to secure the world's most important software. On Jun 2, Anthropic said it's adding about 150 new organizations. Each must meet Anthropic's security requirements before it gets access.

The new group is based in more than 15 countries. It adds fields that weren't well represented in the first group. They include power, water, healthcare, communications and hardware.

In early April, Anthropic said roughly 50 first partners had access to Claude Mythos Preview.

Anthropic estimates a major attack on most partners could affect more than 100 million people.

Anthropic expects many other AI companies to have Mythos-class models in 6 to 12 months. It says they could release them without safeguards against misuse.

Anthropic says it's working to safely release Mythos-level skills in general access. It says the safeguards it needs for that don't exist yet. To its knowledge, no other AI developer has them either.

Anthropic sends the SEC a private IPO draft

Anthropic took a step that gives it the option to go public.

On Jun 1, Anthropic said it sent the US SEC a draft Form S-1, in confidence. It's for a proposed initial public offering of its stock.

Anthropic says this gives it the option to go public after the SEC's review. The IPO will depend on market conditions and other factors. The number of shares and the price aren't set yet.

Less than a week earlier, it raised $65 billion (May 28). Anthropic says that round valued it at $965 billion post-money.

Anthropic says hackers now use AI deeper in their attacks

Anthropic studied a year of banned hacker accounts. It says more of them now rank as higher risk.

Its Jun 3 report covers 832 accounts banned for malicious cyber activity. The bans ran from March 2025 to March 2026.

They're a subset of all bans in that time. Anthropic picked the ones with enough detail to study.

Anthropic scored each actor's risk. In the first six months, 33% ranked medium risk or higher. In the next six months, 56% did.

The most common uses were prep work, like writing malware. Of the 832 accounts, 560 used AI for that.

Per Anthropic, use shifted over the year. It moved from breaking into systems toward work done once inside.

MITRE ATT&CK is a database of attack techniques. Anthropic says it doesn't yet cover AI agents running attacks on their own. It's in talks with MITRE about how the framework might change.

Some results also appear in Verizon's 2026 Data Breach Investigations Report.

Claude had a nearly six-hour incident on Jun 2

Claude had errors on several models early on Jun 2, Central time.

The incident was open from 06:04 to 11:49 UTC. That's 1:04 to 6:49 am Central. It covered claude.ai, the Claude Console, the Claude API and Claude Code.

An Anthropic source spoke to The Register. It said some users "may have experienced intermittent issues or slower response times." Customers on Google Cloud's Vertex AI or Amazon Bedrock were not affected, it said.

On Jun 5, errors hit many Claude models. They started at 15:08 UTC (10:08 am Central). Per the status page, the listed models recovered between 15:25 and 17:29 UTC.

Claude Code now asks before writing files that can run code

Claude Code got more safety prompts, a new trigger word and version limits for orgs.

It shipped versions 2.1.160 to 2.1.168 from Jun 2 to 6, per its changelog.

  • Claude Code now asks before writing some shell startup files, like .zshenv, and git config in ~/.config/git/. The changelog says those writes could otherwise lead to unintended command execution (2.1.160).
  • acceptEdits mode now asks before writing build-tool configs that grant code execution. One is .npmrc (2.1.160).
  • These writes are auto-approved only in bypassPermissions mode, per the week 23 notes.
  • The trigger word for dynamic workflows changed from "workflow" to "ultracode." Asking for a workflow in your own words still works (2.1.160).
  • Organizations can now require an approved version range. Claude Code refuses to start outside it and points users to an approved version. Per the week 23 notes, claude update, install and doctor still work (2.1.163).
  • A failed Bash command no longer cancels other tool calls in the same batch. Each returns its own result (2.1.161).

Anthropic says Opus 4.7 matches or beats NMR software, on average

In a small Anthropic test, Opus 4.7 predicted NMR spectra as well as dedicated software, on average.

Anthropic's Jun 5 post tested Claude on NMR spectra. It compared three Claude models with ChemDraw and MestReNova on 20 compounds. The compounds came from preprints posted after the models' training cutoff.

On hydrogen, Opus 4.7 was the most accurate. On carbon, Opus 4.7 and MestReNova were effectively tied.

Opus 4.7 has no chemistry-specific fine-tuning. Anthropic says it's now as good as or better than both tools on average, for routine data prediction.

Claude also worked in reverse, from spectra and the exact formula to a structure. Each time, it could propose up to three ranked structures. Opus 4.7 got all eight simpler structures on every try.

The seven harder ones came with a hint, the starting material. Opus 4.7 got four right on all three runs, and the rest on two of three. Without the hint, Anthropic says the model could loop without committing to an answer.

Anthropic says the test was small, and the results should be read as indicative rather than precise. 2D NMR and stereochemistry were left out by design, so complex natural products weren't tested. Some scaffolds and solvents went untested.

Quick hits

  • The Claude API stopped billing for refusals with no output (Jun 2). That's when Claude refuses before writing anything, per the release notes.
  • The advisor tool can now cap its output on each call (Jun 2). Anthropic says that cuts latency and output token cost when full-length advice isn't needed.
  • Enterprise plans can now give admin permissions through custom roles (Jun 2). Members can get areas like billing or privacy without being Owners, per the release notes.
  • Anthropic added a three-tier Services Track to the Claude Partner Network (Jun 3). Per Anthropic, more than 40,000 firms have applied to the network since March.
  • The new Claude Partner Hub lists each partner's standing in a public directory (Jun 3). Partners can connect it to Claude through a new MCP connector, per Anthropic.
  • Some Claude Code services were degraded on Jun 3, from 04:17 to 07:36 UTC. That's 11:17 pm Jun 2 to 2:36 am Central. The status page names security reviews, code reviews, routines and some web sessions.
  • Claude Code's claude mcp list, get and add no longer print secrets (2.1.161, Jun 2). Credential headers and URL secrets are redacted.
  • Claude Code messages relayed from other Claude sessions lose user authority (2.1.166, Jun 6). Receivers refuse relayed permission requests, and auto mode blocks them.

Worth your time

That's the week. The next issue lands Sunday, Jun 14.

Reply and tell me what you're building. I read every reply.

— Sam

fuel for the obsessed builder

Subscribe to ClaudeUser

One email every Sunday with everything that changed in Claude this week, with every claim linked to its source.

Free. Unsubscribe anytime.