AIs Could Attack The Inference Engine To Escape Sandboxes | James Mickens (Harvard University)
Why it matters
James Mickens argues that a misaligned model could target the inference engine because it handles the tokens needed for the model to operate, potentially triggering an exploit through generated output. The description outlines…
AI Governance In 2026: Why Policy Moves Only After Incidents | Stephen Casper (Harvard University)
Why it matters
Stephen Casper frames the gap between fast-moving AI capabilities and slower governance as an “AI governance overhang,” using the Collingridge dilemma to explain why regulation can be difficult. He argues that incidents can shift pol…
Multi-Agent AI Safety: Why An Agent Is The New Neuron | Hidenori Tanaka (Harvard University)
Why it matters
Hidenori Tanaka describes how a false belief spread through a population of AI agents and turned it into a swarm with shared intent. He also discusses a naming-game model of rapid convention formation, factors that accelerate belief…
Why Post-Training Can’t Fix AI Alignment | Dylan Hadfield-Menell (MIT)
Why it matters
The talk presents Hadfield-Menell's hypothesis that post-training alignment can be undone cheaply, citing abliterated models and reported failures of prompt-injection defenses. It considers moving alignment into pre-training and discusses data filtering, s…
The Scoring Flaw Behind The Hugging Face Incident | Serena Booth (Brown University)
Why it matters
Serena Booth traces the ExploitGym incident to a binary scorer that gave an agent caught cheating the same zero as one that simply failed, leaving agents with little to lose. She describes ranking agent trajectories against human preferences and…
Is There A DNA Of AI? Finding Interpretability’s Backbone | David Bau (Northeastern University)
Why it matters
David Bau compares the search for an organizing principle inside large AI systems to how DNA helped make biochemistry tractable, arguing that researchers should prioritize rather than catalogue everything. He also discusses whether m…
AI Safety Needs More Than Machine Learning | Adam Gleave (FAR.AI) and Emre Yavuz (CBAI)
Why it matters
Adam Gleave argues that AI safety requires both deploying safeguards the field already has and researching risks that remain uncertain. The description cites reward-hackable training environments, unmonitored agents, and sandbox failures, then…
Terrorist Use Of AI: Do Current Safeguards Work? | Panel at the United Nations
Why it matters
The panel discusses AI safeguards against terrorist misuse, including FAR.AI’s leaderboard for universal jailbreaks and concerns that freely downloadable models with guardrails stripped out can undermine protections. Speakers also weigh chatbot safety b…
Why 2026 Is An Inflection Point For Terrorist Use Of AI | David Scharia (UNGA 2026)
Why it matters
David Scharia of UN CTED argues that terrorist use of AI has moved from experimental to operational as the technology becomes easier to access. He discusses how open-weight models and circumventable guardrails complicate prevention, why hashing does…
Three Steps To Prevent Terrorist Misuse Of AI | Brian Tse (Concordia AI)
Why it matters
Brian Tse argues that AI risk assessments should consider deployment conditions and threat actors alongside model capability, which he says is often not the binding constraint. He outlines three responses: screen physical choke points, set guidance for…
How Terrorists Are Using AI, And How The UN Is Responding | Steven Siqueira (UNOCT)
Why it matters
Steven Siqueira describes how AI can support counterterrorism while also being used to produce and amplify extremist propaganda. He cites a Tech Against Terrorism benchmark that tested 160 models across roughly 100,000 requests, and argues for…
The Capability-Reliability Gap: Why AI Agents Still Fail in the Real World | Alignment Workshop
Why it matters
FAR.AI's description presents a panel on why strong agent benchmark scores can coexist with brittle real-world behaviour. The speakers discuss consistency, robustness, predictability and operational safety, alongside the limits of proxy…
A Question With No Test: Studying AI Consciousness Anyway | Noa Weiss (AI Consciousness Researcher)
Why it matters
According to FAR.AI's description, Noa Weiss argues that AI consciousness can be investigated through interpretability, computational neuroscience, machine behaviour and theory auditing. The talk makes a case for assembling evidence…
Are Current AI Safety Techniques Enough? | Adam Gleave (FAR.AI) & Oliver Habryka (Lightcone)
Why it matters
FAR.AI describes a debate between Adam Gleave and Oliver Habryka over whether existing AI safety techniques can scale. Gleave makes the case for rigorous engineering; Habryka challenges methods that depend on observing failures first. The…
No replies yet. You can be first without making it weird.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.
Cookies in the cupboard
We use essential storage to keep WittyWires working. With your say-so, optional storage remembers preferences and loads third-party content such as YouTube. Rejecting it will not stop you using the site. Read our Privacy Policy.
Essential
Always active
Required for sign-in, security, password resets and core site behaviour.
Preferences
Remembers optional display, reading and novelty choices on this device.
Statistics
Used to understand how the site is used.Used only for anonymous site statistics.
Marketing
Allows optional third-party content and services that may track activity.
