Institute for AI Policy and Strategy (IAPS) retweeted
A key provision in the new AI EO is the vulnerability-discovery "clearinghouse." A good start — but the admin must build on it to meet the defensive-coordination challenges that offensive cyber agents will create. In a just-published @IAPS and @aisafetysg report, we detail how 🧵whitehouse.gov/presidential-…
Institute for AI Policy and Strategy (IAPS) retweeted
While initiatives like Project Glasswing provide tech companies with access to cyber-capable models, "access is only the starting point".
A timely policy memo on AI cyber defence, by @covinstantinop, @JKraprayoon, & @MattInThemittel @iapsAI
Institute for AI Policy and Strategy (IAPS) retweeted
How do you prevent a military coup after the military has been automated? Dave Banerjee (@DaveRBanerjee) of IAPS (@iapsAI) wants Pivotal fellows working with him on AI-driven power concentration. Topics include threat models for how concentration unfolds, and how to govern model specs and constitutions.
Meet Our Mentors ↑
Apply by May 3rd: pivo.tl/fellowship
Fully funded (£6-8k stipend), 9 weeks in London (+ up to 6-month extension), working with senior researchers in AI safety, AI policy, technical governance, and AIxBio.
Institute for AI Policy and Strategy (IAPS) retweeted
I’m very pleased to have had my first policy memo @iapsAI published!
Shortly after the memo was finalised, MiniMax announced it was too offering an ‘always on’ AI agent service, similar to Moonshot.
There are serious security risks associated with these offerings. Depending on how our policymakers respond, they could make the TikTok case in the US look trivial with the benefit of hindsight.
Right now, there is an opportunity to act proactively to manage the security risks associated with these tools and ensure the public is properly informed about the potential implications for their data security and privacy.
As we learnt from the TikTok case, the room for political manoeuvre can be much smaller when the target of government action is something that has become deeply embedded in our lives.
I’m all for having agents in my life but they must be reliable, robust and secure. KimiClaw and MaxClaw are far from that!
Thanks to my wonderful IAPS colleagues for all of their support and feedback on this first piece. Here’s to many more!
📢 New IAPS policy memo: Kimi Claw - Risks from Chinese-Hosted 'Always On' AI Agents.
TikTok collects data from one app. An 'always-on' AI agent can have access across a user's files, apps, and communications all day, every day.
Beijing-based AI company Moonshot is now offering exactly this through Kimi Claw. @theobearman examines the risks and recommends four low-cost steps the federal government can take now before mass adoption risks making intervention harder.
Read the full memo: iaps.ai/research/kimi-claw-r…
Institute for AI Policy and Strategy (IAPS) retweeted
The data risks must be taken very seriously given the unique depth of data these technologies can access. In fact, Koi recently discovered Chinese coding agent extensions are indeed sending *any* file a user opens to servers in China. h/t @theobearman koi.ai/blog/maliciouscorgi-t…
📢 New IAPS policy memo: Kimi Claw - Risks from Chinese-Hosted 'Always On' AI Agents.
TikTok collects data from one app. An 'always-on' AI agent can have access across a user's files, apps, and communications all day, every day.
Beijing-based AI company Moonshot is now offering exactly this through Kimi Claw. @theobearman examines the risks and recommends four low-cost steps the federal government can take now before mass adoption risks making intervention harder.
Read the full memo: iaps.ai/research/kimi-claw-r…
Institute for AI Policy and Strategy (IAPS) retweeted
Adversaries can tamper with or poison leading US models. There also can be risks from insider threats, including potentially the AIs themselves.
@DaveRBanerjee at @iapsAI has a roadmap for how to defend
New report by @onni_aarne and me at @iapsAI 🧵
AI integrity means ensuring AI systems are free from backdoors, poisoned training data, and secret loyalties that could compromise their behavior.
It's one of the most important and least-explored problems in AI security.
Institute for AI Policy and Strategy (IAPS) retweeted
New report by @onni_aarne and me at @iapsAI 🧵
AI integrity means ensuring AI systems are free from backdoors, poisoned training data, and secret loyalties that could compromise their behavior.
It's one of the most important and least-explored problems in AI security.
Institute for AI Policy and Strategy (IAPS) retweeted
Come work on my team at @iapsAI!
We run the (imo excellent) AI Policy Fellowship and are bursting with ideas for more programs—constrained by scaling the team.
There are an Events and a general Programs role open now!
iaps.ai/careers
Institute for AI Policy and Strategy (IAPS) retweeted
New analysis from my colleague Erich Grunewald in The Substrate: Where will China get its compute in 2026?
Four pathways:
(1) ~60% legal NVIDIA H200 imports (~230K B300-eq) - the US could cut this off entirely if it wanted
(2) ~25% domestic Huawei Ascend 910Cs via SMIC (~40K B300-eq) bottlenecked by CXMT's HBM production, not chip fabrication
(3) ~10% smuggled Blackwells (~20K B300-eq) - DeepSeek reportedly already using "several thousand" for next-gen models, according to @theinformation
(4) <5% proxy fabrication via front companies (~2K B300-eq) - the old TSMC stockpile is likely exhausted
Total: ~320K B300-equivalents. Enough for ~6 Grok 4-scale training runs simultaneously and ~600x what DeepSeek reported it used for V3. However, this is far less than what US companies will acquire. The Stargate campus alone will house ~300K B300-eq.
Export controls are the main reason Chinese models lag behind the frontier. To maintain and widen this gap, the US could tighten HBM and semiconductor equipment controls and/or reverse the H200 decision, lower the volume cap, or, at minimum, avoid raising any of the volume cap or performance thresholds.
Outside of this, preparing cloud access restrictions via the Remote Access Security Act, creating a whistleblower incentive program and improving BIS capacity can also be considered to bear down on remote access.
Read the full post here: the-substrate.net/p/where-wi…
CC @ohlennart
Institute for AI Policy and Strategy (IAPS) retweeted
Its wild that after China launched the first end-to-end AI automated cyber campaign, we are still supplying 60% of the tools they need to go even further. AI chips = cyber capabilities. We should treat them as such. h/t Erich Grunewald. the-substrate.net/p/where-wi…
Institute for AI Policy and Strategy (IAPS) retweeted
What future do we want for advanced AI? This depends on your "strategic vision" – your high-level view on navigating the transition to superintelligence. Our new @iapsAI report maps 9 distinct visions from status quo to global government project 🧵
Institute for AI Policy and Strategy (IAPS) retweeted
SEC lets whistleblowers report financial fraud. Could we do the same with export controls? Congress is considering the "Stop Stealing Our Chips Act" to do exactly this. Erich covers it in a new blog post.
Replying to @iapsAI
Second, Erich on a new bill that'd create an SEC-style whistleblower program for export violations by rewarding folks who report violations to US authorities with 10-30% of any resulting penalty. He argues this bill could dramatically improve enforcement while paying for itself.
the-substrate.net/p/the-case…
Institute for AI Policy and Strategy (IAPS) retweeted
The US is selling H200s to China. More advanced than what China can manufacture, which means we boost Chinese AI.
But @onni_aarne is right - we pay too much attention to the quality of the chips... and not about the *quantity*! Nvidia manufactures a lot more than China can.
Replying to @iapsAI
The Substrate is launching with two substantive posts.
First, @onni_aarne argues that export controls focus too much on AI chip quality and not enough on quantity. This has significant implications, not least for how to think about the recent NVIDIA H200 decision.
the-substrate.net/p/for-chip…
Institute for AI Policy and Strategy (IAPS) retweeted
New wonky Substack just dropped on AI compute policy -- called The Substrate, and run by the @iapsAI compute policy team. So excited!
Institute for AI Policy and Strategy (IAPS) retweeted
The IAPS compute policy team has launched a newsletter, The Substrate, and I just published my first post! “For chip exports, quantity is at least as important as quality”
Institute for AI Policy and Strategy (IAPS) retweeted
IAPS has a new brief explaining the recent AI-enabled cyberattack reported by Anthropic.
Institute for AI Policy and Strategy (IAPS) retweeted
One way the US could lose the AI race: if companies fail to secure their most advanced, non-public models from theft or sabotage.
Here’s how to secure AI’s hidden frontier👇 w/@iapsAI
Institute for AI Policy and Strategy (IAPS) retweeted
Job Opportunity!
The Institute for AI Policy and Strategy seeks a Senior Research Manager/ Director of Research - AI Frontier Security. @iapsAI
daybook.com/jobs/Br8qbwcDqdh…
Institute for AI Policy and Strategy (IAPS) retweeted
AI has "the potential to reshape the global balance of power, spark entirely new industries, and revolutionize the way we live"
This isn't hype. It's official US policy.
And with bipartisan members of Congress giving AI timelines, we're seeing a vibe shift. I unpack it here:
Institute for AI Policy and Strategy (IAPS) retweeted
What should we do about chip smuggling?
A new @iapsAI @CNAStech report outlines policy:
- Whistleblowing program to expand enforcement
- Chip notification program to give enforcers vital data (currently they fly blind)
- partnering with the US intelligence community






























