AI Risk Atlas Prototype/Demo — Unofficial independent experiment. Not an official xAI product. Scores can be wrong.
Unofficial capital desk
Public-reported rounds and grants that name AI safety or security as the product. Amounts are compiled estimates from company posts and press — not cap tables, not official, and not a recommendation.
Scores are automated, experimental estimates from public X posts and a hand-written seed corpus. They are not formal risk assessments, not certified, and not suitable for compliance or operational decisions.
Consequence, likelihood, and urgency are 1–5 judgements applied by this project, not by a standards body. Residual scores assume only the mitigations marked in place. A signed-in reviewer can override residual and mark an item reviewed — that override is still unofficial. Aspect tags (capability, domain knowledge, affordance, impact domain) are a lightweight PRA aid, not a formal hazard analysis.
Each flow is split evenly across the risks it names. That is a bookkeeping choice, not a true allocation.
Interpretability lab · as of 2026-02 · venture
Decode and steer model internals so deception and goal-drift are inspectable.
Feb 2026 Series B $150M at $1.25B; prior ~$57M. Interpretability is the control, not a guarantee.
Targets · Deceptive agent behavior in the wild · Loss of human agency through over-delegation · Refusal collapse under long chain-of-thought · De facto clinical AI without clinical controls
Source · Goodfire Series B postRed-team / agent security · as of 2026-05 · venture
Adversarial testing and runtime protection for frontier and enterprise agents.
Cited on 11 frontier system cards. Office expansion reported 13 Aug 2026.
Targets · Prompt injection of institutional systems · Refusal collapse under long chain-of-thought · Deceptive agent behavior in the wild · AI-authored software vulnerabilities · Agent skills leaking live credentials
Source · Gray Swan Series A (28 May 2026)Model security platform · as of 2025-2026 · venture
Detect attacks on models in production — extraction, inversion, prompt abuse.
Amount is a public-report midpoint, not a filing.
Targets · AI-authored software vulnerabilities · Agent skills leaking live credentials · Prompt injection of institutional systems · Ungoverned open-weight proliferation
Source · Field compilation (2026)Prompt / runtime guard · as of 2025-2026 · venture
Stop jailbreaks and injection before they reach the model.
Guardrail vendors address a slice of injection, not agentic autonomy.
Targets · Prompt injection of institutional systems · Refusal collapse under long chain-of-thought · Agent skills leaking live credentials
Source · Field compilation (2026)Technical AI safety RFP · as of 2025-02 · grant
Misalignment research — evals, oversight, control, academic labs.
Typical grants $50k–$5M. This is one of the few large dedicated safety pots that is not a lab raise.
Targets · Deceptive agent behavior in the wild · Sandbox and containment escape · Frontier safety governance rollback · Loss of human agency through over-delegation
Source · Coefficient Giving RFPAISI-hosted coalition grants · as of 2026-02 · grant
60 alignment projects — evals, oversight, control, deception tests.
Coalition money, not a VC round. OpenAI and Microsoft are both funders and evaluatees.
Targets · Deceptive agent behavior in the wild · Sandbox and containment escape · Frontier safety governance rollback
Source · AISI Alignment Project (19 Feb 2026)Quarterly philanthropy platform · as of 2026-08 · grant
Field capacity — people, orgs, and infrastructure around catastrophic risk.
Round was still open at last sweep. Treat as intended, not landed.
Targets · Frontier safety governance rollback · Closure of the entry-level labor market · Loss of human agency through over-delegation
Source · AISafety.com funding deskLab safety grants · as of 2026-01 · grant
Independent research on companion / wellbeing harms.
Closed Jan 2026 after 1,000+ applications. Tiny next to companion-product revenue.
Targets · Companion models and harm to minors · Loss of human agency through over-delegation
Source · OpenAI grant noteThese raises dwarf the safety stack. We keep them separate so a $5B NVIDIA check is not counted as residual reduction.
Build a safe superintelligence — compute and talent first. Safety is the product claim, not a side lab.
Most of this is training compute. Treat only a slice as safety residual reduction.
Source · Reuters / Seedtable compilationSeabed mapping and marine infrastructure perception — capability into the water column, not a safety lab.
Capability capital aimed at maritime perception. Do not count as residual reduction on MASS/GNSS risk unless the product ships spoof-detection as a class item.
Source · The SaaS News seed note