postgrok

anthropic

Properties3
tagsorg, ai-lab
typeentity
statusseedling

2 min read

Anthropic

The lab behind Claude and Claude Code. Two facts the garden’s pages depend on, both verified against primaries.

First, the bun acquisition, announced 2025-12-03, positioning Bun as infrastructure for Claude Code and later AI coding products https://www.anthropic.com/news/anthropic-acquires-bun-as-claude-code-reaches-usd1b-milestone. From it came the 2026 Rust port: jarred-sumner, at Anthropic, migrated Bun from Zig to Rust with Claude agents, test-anchored, roughly 535,000 lines of Zig, 19 post-merge regressions all fixed, 128 open bugs closed https://bun.com/blog/bun-in-rust. Anthropic’s own migration guide cites it as the flagship case for large-scale code migrations with Claude Code https://claude.com/blog/ai-code-migration. The whole chain is the empirical support behind rust-for-agents.

Second, the 2025-10-09 poisoning research with the UK AI Security Institute: around 250 poisoned documents sufficed to implant a backdoor across models from 600 million to 13 billion parameters, with attack success depending on the absolute count of poison documents rather than the percentage of training data, and no effective mitigations reported at the time https://www.anthropic.com/research/small-samples-poison. See local-ai-threat-layers. Note: 2026-10-21-youtube-kai-local-ai-privacy also credits “Anthropic 2024” with the finding that backdoors survive safety fine-tuning; the verified Anthropic result in this ingest is the 2025-10 study above, and the 2024 attribution stays flagged.

Third, thinner: 2026-10-12-youtube-indydevdan-pi-durable claims the big labs build products on durable agent substrate reachable from phone, desktop, and app. Unverified, recorded on the durable-agents page, not here.

Sources

Graph View

Backlinks