Attackers only need 1 weak link
AI is getting good at finding software flaws. Defenders have to patch every one; an attacker only needs a single miss.
Need a way to prove code has no flaws.
Potentialintervention
Potentialintervention
Potentialintervention
ForecastedAI risk
ForecastedAI risk
ForecastedAI risk
Current AIprogress
Some of AI’s biggestrisks are no one’sresponsibility
AI capabilities are pushing society into uncharted territory, creating risks whose solutions span companies, labs, funders, and governments — and so fall to none of them.
Atlas makes them someone’s job
We pick out the unowned risks that matter most, design a credible intervention for each, and empower the person or institution best positioned to carry it forward.
The risks arealready visible.The path to asolution is not.
These are not abstract concerns from hypothetical thought experiments. They are visible cracks in society’s resilience posture that have been predicted to grow by the best forecasters.
Projects
Each problem is foreseeable.
Each may be preventable.
But only if someone is responsible for the fix.
Humanity has no say
The behavior of leading-edge AI is determined by the decisions of a handful of leaders at an AI lab. The rest of humanity has no path to weigh in, and it’s no one’s job to build one.
Need an outside convener the labs will actually join.
Unverified
Unverified
You can’t tell what’s real
An AI enabled disinformation campaign cheaply floods news feeds with text, images, and video arguing a crisis both is and isn’t unfolding.
Need shared provenance people can rely on.
Most institutions are built around the approach.Atlas is built around the outcome.
| Role | Approach | Outcome |
|---|---|---|
| Investor | fund an ambitious founder + idea | a product with market demand |
| Researcher | pursue an open question | published results |
| Policymaker | act on a mandate | a new law or policy |
| Think tank | analyze a problem for society | proposed solutions |
| Standards body | convene an industry | consensus around a standard |
| Technical recruiter | fill a defined role | a qualified hire |
   Atlas | whatever maximizes impact — a new org, company, coalition, policy, or contract | the outcome society needs |
Atlas isn’t a think tank, a technical recruiter, or an investor. Each of those is optimized for a niche, with deep expertise built around a proven methodology; Atlas starts with the problem, then chooses the approach that best fits.
Alreadylaunched
AI-driven Formal Verification
Atlas identified that no approaches to formal verification were heavily betting on AI systems improving, and made the case to Jason Gross 
Jason Gross Theorem Labs LinkedIn that the bet was worth making. [read about the work]
Today, Jason leads Theorem Labs, pioneering AI-first approaches to formally verifying software.
Tamper-Response for Compute Accelerators
Atlas identified the need for tamper-response systems to protect AI hardware, and made the case to Mehmet Sencan 
Mehmet Sencan CEO, EARENDIL LinkedIn that he was the one to build them.
Today, Mehmet leads Earendil, which develops scalable, safe backstops for AI compute security.
Scalable Oversight via Formal Specifications
Atlas identified specification validation as a critical bottleneck for scalable oversight of AI, and made the case to Mike Dodds 
Mike Dodds Oath Technologies LinkedIn that better tools were worth building.
Today, Mike runs Oath Technologies, a Focused Research Organization building the specification tools and processes that make scalable oversight of AI possible.
GetInvolved
Atlas is a non-profit organization. Send a risk, bring your skills, or fund the work.
Fund anExploration
Some funders come to Atlas with a concern: cyber-resilience, information integrity, or another AI risk no one clearly owns. Atlas maps concern to a concrete intervention: scoped, tested, and ready for the right owner.
Bring YourSkills
Apply to join Atlas full-time as a Field Strategist or support an existing effort. Atlas needs people who can scope problems, pressure-test plans, and convene the stakeholders who make a fix real.
Send aNeglectedAI Risk
Tell us about a society-scale AI risk where the owner is missing. What’s the clearly desirable outcome that needs Atlas to find a path forward?