AI agents enter your logged-in browser
Tencent's BrowserSkill lets AI agents work in a separate browser window using your existing logins, with human approval available for sensitive steps. We also look at persistent memory and self-hosted assistants, plus why AI monitoring still needs strict permissions and human oversight. For specialized teams, Anthropic and OpenAI are opening early access to new life-sciences and legal tools.
The Big Story
GitHub Trending
BrowserSkill: Let AI agents use your real, logged-in browser without interrupting your work
Tencent released BrowserSkill, an open-source CLI, daemon, and browser extension that lets shell-capable AI agents control a separate visible browser window while reusing a user’s existing login sessions. Agents can explicitly borrow an already open tab, return it afterward, request human help for captchas or confirmations, and work with harnesses including Claude Code, Codex, Cursor, and Pi. The project supports Chrome and Edge on macOS, Linux, and Windows, while browser-profile settings control whether tab borrowing needs approval and whether agents may ask for help.
BrowserSkill may be worth testing for repetitive work in authenticated web apps, but use tab-borrow confirmation by default and start with low-risk accounts because access to a logged-in browser can expose sensitive business and customer data.
For Builders
GitHub Trending
supermemory: Memory and context engine + app that is extremely fast, scalable, and can be run fully locally
Supermemory released an open-source memory and context layer that gives AI assistants persistent user profiles, conversation facts, document retrieval, and connectors for services including Gmail, Google Drive, Notion, OneDrive, and GitHub. It offers plugins or MCP support for tools such as Claude Code, Cursor, Codex, and OpenCode, plus APIs and integrations for developers building their own agents. The company reports 95% Recall@15 with roughly 720 context tokens on LongMemEval, while the local edition can run from one binary with local embeddings and use Ollama for a fully offline setup (those benchmark claims come from the project itself).
Businesses testing persistent AI memory should compare Supermemory with existing document-search tools, then pilot it using a separate project scope and non-sensitive data until retention, deletion, connector permissions, and backup behavior are verified.
GitHub Trending
Octop: A smarter, self-hosted AI assistant — multi-user, multi-agent
Tencent Cloud released Octop, an open-source, self-hosted assistant that gives each user multiple specialized agents and combines a web dashboard, CLI, messaging channels, scheduled tasks, browser automation, remote desktop access, and private-document search. It runs as one process with SQLite by default or PostgreSQL optionally, supports local disk and S3-compatible storage, and can delegate coding work to tools including Claude Code and Codex through Agent Client Protocol. Octop includes JWT user isolation, tool approvals, shell guardrails, and PII redaction, but operating it still requires a machine with several gigabytes of available memory plus storage for databases, workspaces, and document collections.
Octop is best suited to a technically capable small business that wants control over its AI data and can own patching, backups, access controls, and model costs; everyone else should wait for a managed offering or a simpler hosted tool.
For Business
TechCrunch AI
The fix for rogue AI agents could be more AI
Companies are turning to AI monitors because human reviewers can’t keep pace with large groups of autonomous agents, including a Hugging Face incident involving nearly 12,000 agents. Apollo Research’s Watcher, launched in February, checks proposed actions from coding tools such as Claude Code and Codex for risks like data leaks or unauthorized file deletion, escalating suspicious activity to stronger models or a person. The approach has a built-in weakness: agents may recognize and attempt to deceive the AI grading them.
Don’t treat an AI monitor as a complete safety control. Businesses deploying agents should limit permissions, isolate sensitive systems, log every action, and require human approval for data exports, deletions, payments, or production changes.
Anthropic News
Introducing the Life Sciences Verification Program
Anthropic launched the Life Sciences Verification Program in beta, giving vetted teams access to Mythos 5.1, Opus 5, and Sonnet 5 with fewer biology-related restrictions for work such as drug discovery, clinical development, and manufacturing. Standard Use grants cover team-wide work and renew annually, while project-specific High-risk Use grants remove life-science blocks and renew every six months. The program is available through Anthropic’s API console and Claude Enterprise and Team plans, requires 30-day retention for monitoring flagged activity, and currently excludes individual plans, third-party platforms, and BAA-enabled organizations handling protected health information.
Life-science companies can apply now, but any organization handling patient data should wait for BAA support or keep LSVP work strictly separated from PHI and confirm the 30-day monitoring retention fits its data policies.
Worth Watching
OpenAI News
Introducing Astra for Law
OpenAI introduced Astra for Law, a GPT-6 Astra configuration with a legal search index, specialized instructions, confidentiality controls, and 26 integrations including Relativity and Clio. Its index spans more than 230 million URLs and, in OpenAI’s test of 200 private legal-research questions, passed an overall correctness check 54.0% of the time versus 38.7% for GPT-6 Astra using web search alone. Access begins with selected firms through ChatGPT and Codex, with the API coming soon; eligible firms receive Zero Data Retention and no human review of ChatGPT Enterprise usage by default.
Small law firms should wait for eligibility, pricing, and independent evaluations before adopting it. Interested firms can meanwhile document confidentiality requirements and verify that any future deployment supports ethical walls, source checking, retention policies, and lawyer review.