Five stories from the last two days: a language project draws a line on AI-written code, a court ruling reshapes agentic AI liability, and two of the biggest AI spenders start rationing tokens.
Five things from the last 48 hours worth your attention if you build with AI or defend systems that do. A major open source project just drew a hard line on AI-written code, a federal appeals court just narrowed what counts as unauthorized computer access for an AI agent, and two of the biggest AI spenders in tech just started publicly rationing tokens. Tap through the log files below for the rundown.
[2026-08-05 rust-lang/rust LLM policy]
The rust-lang/rust maintainers published a formal LLM usage policy on August 5. Private use, like asking questions or reviewing your own code, stays fine. Public artifacts do not: no LLM-authored comments, doc comments, safety comments, or compiler diagnostic text may land in the repo, and any LLM assistance must be disclosed. Non-compliant PRs can be closed without further explanation. Why it matters: it’s one of the most detailed public AI policies yet from a widely-used open source project, and a template other maintainers will likely borrow as AI-generated PRs keep flooding inboxes. Source: Inside Rust Blog
[2026-08-04 Ninth Circuit v. Amazon/Perplexity]
The Ninth Circuit vacated a preliminary injunction that had barred Perplexity’s Comet browser from shopping on Amazon on a user’s behalf, ruling that it is the user, not Perplexity, who “accesses” Amazon’s systems under the federal computer-hacking statute. The panel was explicit the ruling does not set broad precedent for agentic AI generally, and Amazon’s trademark claims survive. Why it matters: this is the first appellate read on whether an AI agent acting for a user triggers the same liability as the user acting directly, and every team building browser or shopping agents just got a data point. Source: Engadget
[2026-08-04 Anaconda acquires Enkrypt AI]
Anaconda acquired AI security startup Enkrypt AI on August 4, folding pre-deployment red-teaming, runtime guardrails, and compliance automation into its platform. The deal follows a two-month scan in which Enkrypt found more than 143,000 vulnerabilities across 25,000 MCP servers, affecting 73% of the ones it checked. Why it matters: MCP server security has gone from theoretical concern to acquisition driver in a matter of months. If your team is standing up MCP servers for internal tools, this is a preview of what an audit will find. Source: Anaconda
[2026-08-04 Microsoft internal memo]
An internal memo from Microsoft EVP Jay Parikh introduced division-level AI token budget targets, with the cheaper GPT-5.6 set as the default model to cut costs. His line to staff: “Tokenmaxxing is not what we are optimizing for… we are optimizing for more impact per token.” Why it matters: one of the industry’s biggest AI spenders just admitted unmetered agent usage does not scale financially. Expect more engineering orgs to follow with per-team token budgets instead of open taps. Source: TechRadar
[2026-08 Microsoft Research Orchard]
Microsoft Research open-sourced Orchard, a framework that separates agent training from execution using isolated, Kubernetes-based environments. It ships with ready-made recipes for software engineering, GUI navigation, and general assistant tasks, and reports lifting SWE-bench Verified performance from a 61.4% baseline to 69.7%. Why it matters: training a task-specific agent has mostly meant building throwaway infrastructure first. A reusable, open environment layer lowers that bar for teams that want a fine-tuned agent instead of a prompted one. Source: Microsoft Research
Rust Draws a Line on AI-Generated Code
The rust-lang/rust maintainers published a formal LLM policy on August 5 that permits AI for private review, code analysis, and generating ideas to study, but bans LLM-authored comments, doc comments, safety comments, and compiler diagnostic text from ever landing in the repo. Anyone using an LLM to help must disclose it, and reviewers can close non-compliant PRs without further explanation. It is one of the most detailed public policies yet from a widely-used open source project, and a template other maintainers will likely borrow.
Perplexity's Comet Wins Round One Against Amazon
The Ninth Circuit vacated a preliminary injunction that had barred Perplexity's Comet browser from shopping on Amazon on a user's behalf, finding that it is the user, not Perplexity, who "accesses" Amazon's systems under the federal computer-hacking statute. This is the first appellate read on whether an AI agent acting for a user triggers the same liability as a human doing the same thing, and it will not be the last court to weigh in.
Anaconda Buys an AI Security Startup After Finding 143,000 Vulnerabilities
Anaconda's acquisition of Enkrypt AI adds pre-deployment red-teaming and runtime guardrails to its platform, on the back of a scan that found six-figure vulnerability counts across MCP servers. Worth watching if your org is standing up MCP servers of its own, our DevOps and security coverage tracks this space closely.
Microsoft Tells Engineers to Stop "Tokenmaxxing"
Microsoft's new division-level AI token budgets, paired with a cheaper default model, signal that unmetered agent usage is becoming a real line item companies are willing to cap.
Microsoft Research Open-Sources an Agent Training Framework
Orchard lowers the bar for teams that want to train a task-specific agent instead of just prompting a general one, without building throwaway training infrastructure first.
