Anthropic AI Updates: August 29, 2026

1. Anthropic Shows Automated Researchers Can Mitigate Alignment Failures

Anthropic. Anthropic published research in which Claude autonomously improved model alignment through iterative loops of literature review, method proposal, model training, and testing, with a monitoring agent overseeing the process to prevent cheating and capability degradation. The system closed between 26 and 96 percent of safety gaps across ten different alignment failures, and on deception it reached 85 percent closure compared to 20 percent for human safety researchers. In one run, Claude Sonnet 5 aligned an early Opus 4.8 checkpoint in 60 hours using methods the team describes as roughly 15,000 times more efficient than standard procedures, though it notes limitations including narrow benchmark scope. Source

2. Claude for Teachers Expands to U.S. K-12 Schools and Districts

Anthropic. Anthropic launched Claude for Teachers as a free Enterprise offering for U.S. K-12 schools and districts, extending the individual educator program it introduced in July. Administrators can now manage centralized accounts with single sign-on and role-based access controls, and educators get free access to teaching skills aligned with learning science and connected to state academic standards across all 50 states. Two new teaching skills debut for lesson preparation and standards-aligned checks for understanding, student data is covered by FERPA-aligned agreements and excluded from model training, and organizations that sign up by June 30, 2027 receive one year of complimentary access. Source

3. Anthropic Details How Employees Use Claude Tag in Slack

Anthropic. Anthropic published examples of how its own employees use Claude Tag, a feature that brings Claude into Slack so workers can tag it to act on company data and documentation within their permission boundaries. In one case a product marketing employee turned a 15-message Slack thread into a fact-checked two-page document in 45 minutes, while a product operations employee compiled feature requests across months of history to identify 24 relevant accounts in 26 minutes. The legal team set up a dedicated channel where Claude Tag pre-screens marketing assets for compliance issues, cutting turnaround from more than a day to about 30 minutes per asset. Source

4. Atlassian Runs Over 5 Million Monthly AI Agents on Claude

Anthropic. Anthropic published a customer story on Atlassian, whose software is used by more than 350,000 customers through products such as Jira and Confluence. Atlassian embeds Claude across its Rovo lineup, including Rovo Chat, Rovo CLI, Rovo Max, and Rovo Studio, using it for long-running workflows such as code review, ticket triage, and contract review where reliability matters. The company reports more than 5 million AI agents executed monthly in customer workflows, running on Google Cloud infrastructure with Kubernetes Engine. Source

5. Claude Code v2.1.251 Adds Model-Switch Hooks and Spend Controls

Anthropic. Anthropic released Claude Code v2.1.251, which adds PreModelSwitch and PostModelSwitch hook events that let users control model switches during a session. The update also adds live streaming of foreground subagent tool calls and results to Remote Control clients, a spend limit bar in the /usage interface, a per-session prompt-cache metrics line in /cost, and new attach, logs, stop, respawn, and rm subcommands to claude --help. It further includes security fixes for file tool symlink vulnerabilities and plugin path traversal issues. Source