| | Making sense of… | | The massive change in how software is produced | | AI coding tools started as helpful autocomplete only a couple of years ago and have evolved beyond that into an impressive new range of task automation. The more capable systems now participate in the full software lifecycle, including reading requirements, writing code, running tests, fixing bugs and iterating on failures. That is incredibly useful as much as it is a new source of risk. Many software engineers and vendors already use these tools daily. In some cases, software systems are 80% written by AI. How to provide adequate system assurance and governance that keeps up with this pace remains an open question. | | Read the full analysis → | |
|
| | On the radar | | 1. Anthropic releases two frontier models followed by sovereign access controls demanded by the US government. Anthropic’s June 2026 release of Claude Fable 5 (public, with classifiers routing sensitive queries to a weaker model) and Mythos 5 (restricted to vetted organisations, classifiers lifted) formalises a new access structure for dual-use AI capability (i.e., AI that could be used for civilian or military/hostile use). Five days after release, Fable was required to be withdrawn for non-US nationals under a US government directive (leading to Anthropic withdrawing it from all users). Two risks stand out for Australian agencies and critical infrastructure operators. First, frontier dual-use capability now sits in a public product, widening the pool of actors (including hostile ones) able to turn frontier tools against the systems these operators defend. Second, access to such capability can be withdrawn without warning by a foreign government, as the five-day Fable episode showed. Both point to the need to manage a growing dependence on frontier models controlled offshore, including continuity plans for when access is restricted. | | 2. “Sequent” launches. Former UK AI Safety Institute researchers led by Geoffrey Irving have merged with AI alignment research organisation Timaeus (co-led by the Australian mathematician Daniel Murfet) to form Sequent, a large non-profit research organisation focused on AI alignment. Credible researchers are concluding that the problem requires dedicated institutional capacity outside frontier labs and are willing to say publicly that they may find only obstacles to alignment being achieved. | | 3. Australia’s AI Safety Institute moves from announced to operational, and Kate Conroy has been appointed as its head. The institute will monitor, test and analyse advanced AI capabilities, risks, harms and trends to support regulators and agencies in responding to these. It will also engage with other AI safety bodies internationally. | | Trendline. The task horizon is lengthening. METR’s March 2025 paper found the length of tasks AI agents can reliably complete has been doubling roughly every seven months. Anthropic’s own data from June 2026 suggests four months. Whether the true figure is closer to four or seven, the direction is unambiguous: systems that could handle minutes of expert-equivalent autonomous work a year ago can now handle hours. |
| |
|
| | Question of the month | | As AI agents take on more of the software lifecycle, what does meaningful human oversight look like and who is accountable when it fails? | | "Human in the loop" has become a phrase that means almost everything and therefore almost nothing. In fact, the research literature distinguishes at least eight variants, each implying a different distribution of decision authority and risk. We’re curious to hear your experiences and takes, we read every answer. | | Reply to share your answer → | |
|
Media and speaker enquiries |
|
|
