Import AI 453: Breaking AI agents; MirrorCode; and ten views on gradual disempowerment
The agent-security and timeline-acceleration threads are directly relevant to APS agencies beginning to deploy or govern agentic AI systems.
Key points
- Import AI issue 453 covers AI coding capabilities, agent security vulnerabilities, policy frameworks, and AI timeline forecasts.
- Google DeepMind's taxonomy of six AI agent attack genres has direct implications for agencies deploying agentic AI tools.
- A curated newsletter rather than a single-issue article; each thread warrants separate follow-up at source.
Implications for Australian agencies
- Consider Agencies developing AI agent use cases or governance frameworks could consider reviewing the Google DeepMind agent-attack taxonomy as an input to threat modelling and assurance documentation.
- Monitor Policy and strategy teams may want to monitor AI capability timeline forecasts, as accelerating progress assumptions affect the planning horizon for APS AI governance frameworks.
Implications are AI-generated. Starting points, not advice — see methodology for how they're framed.
View original source
Copied.
Appeared in:
Weekly digest, 13 April 2026
"Import AI 453: Breaking AI agents; MirrorCode; and ten views on gradual disempowerment"
Source: Import AI – Substack (Jack Clark)
Published: 13 April 2026
URL: https://importai.substack.com/p/import-ai-453-breaking-ai-agents
Jack Clark's Import AI issue 453 covers five distinct developments: the MirrorCode benchmark demonstrating AI can autonomously reimplement complex multi-thousand-line codebases; the Windfall Policy Atlas cataloguing 48 policy responses to transformative AI disruption; a Google DeepMind paper taxonomising six genres of attack against AI agents with proposed mitigations; an AI forecaster doubling the estimated probability of full AI R&D automation by 2028; and a conceptual summary of ten framings for 'gradual disempowerment'. The agent-security paper is the most operationally relevant item for APS practitioners, as agencies increasingly consider agentic AI deployments where prompt injection, behavioural control, and ecosystem-level attacks pose real governance and assurance challenges.
Implications for Australian agencies:
- [Consider] Agencies developing AI agent use cases or governance frameworks could consider reviewing the Google DeepMind agent-attack taxonomy as an input to threat modelling and assurance documentation.
- [Monitor] Policy and strategy teams may want to monitor AI capability timeline forecasts, as accelerating progress assumptions affect the planning horizon for APS AI governance frameworks.
Retrieved from SIMS, 18 July 2026.