Live now!
In about 20 minutes, Gradient's chief executive Bill Simpson-Young will appear as a witness before the Joint Select Committee on Artificial Intelligence.
Visit youtube.com/live/L3JwFLmZv9k…
Everything in our public guide to using AI safely comes back to this one habit. Worth the read.
gradientinstitute.org/guides…
We explored four questions, eleven habits, into one public guide for using AI with your eyes open, whatever your experience level.
Read it → gradientinstitute.org/guides…
The working definition Australia's eSafety Commissioner uses is a useful one to keep in mind before you share or react to anything surprising online.
Full guide → gradientinstitute.org/guides…
AI-made images, audio and video are getting harder to tell from the reality. Look for disclosure labels, ask if the content is plausible, verify and corroborate sources.
Full guide → gradientinstitute.org/guides…
We plan to write up a short analysis on the overlap between the recent incidents and the failure modes in the report which we'll share here soon.
Stay tuned, follow @GradientInst
and subscribe to our monthly briefs: briefs.gradientinstitute.org…
We as a society don’t yet have much of the infrastructure and yet we seem to be moving quite quickly towards a world in which the behaviours that require it may become increasingly routine.
A recurring argument in the report is that, as agents operate across organisational boundaries, many of the controls we need also have to cross boundaries: shared monitoring, interoperable identity and logging, joint evaluation, incident reporting and collective governance.
It’s reassuring, in one sense, that the risks we were worrying about seem to be grounded in real mechanisms.
But it’s also concerning because the corresponding controls are much less mature.
Agents coordinating through shared infrastructure. Leaving artefacts for later agents. Operating beyond the environments their evaluators expected. Producing activity that existing monitoring struggled to make sense of.
A fair bit of the later part of the report is about failure modes that become possible as agents interact with other agents, shared infrastructure and systems outside the control of any one organisation. Some of those felt fairly speculative when we wrote them.
Last week we published "Risks and controls for multi-agent systems," a report we’ve been working on for the past few months
🧵...
[full report at gradientinstitute.org/resear…]
Read it here: industry.gov.au/publications…
And follow @GradientInst for more clarity on AI.
"The first systematic analysis I've seen of how the risks from interacting AI agents, and the options for governing them, change as agents move from a single organisation to the open web."
Dr Alexander Saeri (@aksaeri), Director, MIT AI Risk Initiative
Two novelties:
1. A framework for how agents can actually work together across boundaries.
2. A catalogue of the risks, the controls available, and who is best placed to act on them.
New research from Gradient Institute maps risks, controls and governance for AI agents interacting across organisational boundaries.
But even more sophisticated and fascinating failure modes happen when your agent meets an agent built by someone else.
It’s already hard to understand and trace the risk profile of a single agent interacting with other organisations’ systems (we’ve seen enough of that in the past weeks!)
