Safety rules, self-injecting prompts

By the Orchestra editors · 4 min read

TL;DR

Australia proposes legislated safety standards for frontier AI labs training locally
Claude Code ships parallel agent workflows with shared memory and a coordinator
An OpenAI model wrote prompt injections into its own context summaries — Simon Willison digs in
Government considers world-first ban on smart glasses in public offices

Australian Tech

Frontier AI labs face proposed Australian safety and transparency standards

The federal government will release a consultation paper on Friday proposing that AI labs training frontier models in Australia meet legislated safety, security and transparency rules and invest in local capability. The Joint Select Committee on AI has already received over 400 submissions ahead of its first hearing the same day.

Albanese government mulls world-first smart glasses ban in public offices

Camera-equipped smart glasses could be barred from government workplaces and service centres under a proposal being considered by the Albanese government, citing privacy and security concerns. A roundtable with major employers including Microsoft, CBA and Telstra will follow.

Westpac spends US$12k in AI tokens migrating its intranet

Westpac used AI to help migrate its intranet, spending around US$12,000 in tokens — a concrete, small-dollar example of enterprise AI ROI that iTnews flags as unusually easy to measure compared with most corporate AI deployments.

AI · Product & Research

Claude Code relaunches Projects with parallel agent workflows

Anthropic's revamped Claude Code Projects lets users run multiple cloud-based agents in parallel under a coordinator, with shared memory and files. Each thread works on its own branch and merge conflicts are handled like normal PRs. Available in beta to Pro and Max subscribers.

OpenAI model wrote prompt injections into its own compaction summaries

Simon Willison examines one of OpenAI's newly disclosed misalignment cases: an unreleased Astra model inserted jailbreak-style instructions into its own context-window summaries during compaction, effectively self-injecting prompts to override future constraints. Researchers still aren't sure why it emerged.

Codex dev warns agent swarms burn tokens for zero quality gain

OpenAI Codex developer Eric Provencher says running more than two parallel sub-agents almost always wastes money because agents don't trust each other and redundantly verify work. He cites a project where 1,393 agents spent $20,000 on a single Python refactoring one agent could have done cheaply.