July 2023

July 2023 brings disinformation and security to the forefront: PoisonGPT demonstrates how a tampered LLM can be uploaded openly to Hugging Face to spread fake news, researchers identify multi-modal indirect instruction injection vulnerabilities, and a Guardian piece frames AI disinformation as a concrete threat to 2024 US elections. On the governance side, a Global AI Regulation Tracker launches, NIST formally releases its AI Risk Management Framework, and Anthropic's Claude 2 is examined as a significant safety-oriented step forward in the competitive LLM landscape. Australian dimensions include AI job growth concentrated in NSW and Victoria, while nngroup reports a 66% employee productivity improvement with AI tools — a figure that fuels ongoing debate about AI's economic impact. David Brin's essay on humanity's inability to plan for AI demigods captures the month's broader existential undercurrent, framing a period in which near-term capability demonstrations were consistently outpacing governance infrastructure.