Research hub · started 29 Aug 2026
Can you build a leading defensive AI cyber lab?
A ground-up map of the AI cybersecurity space — the companies, the models, the benchmarks, the money, the law — assembled to answer one question: is there a defensible position for a new lab that makes AI models good at defending code, and if so, what is the first thing to build?
$1B+
XBOW valuation, Mar 2026 — the category's price tag
86 / 68
% of AIxCC synthetic vulns found / patched by autonomous systems
0
C-language real-world bugs patched at AIxCC finals
~56%
of AI-generated code passing security checks (Veracode), flat YoY
<5%
of curl bug-bounty reports now valid, down from >15% (AI slop)
4–7 mo
open-weight lag behind frontier on cyber, per UK AISI
Every figure above is sourced on its page. Figures that did not survive fact-checking were removed and are recorded in the verification ledger.
Reading paths
This is a wiki, not a report. Enter wherever your question is.
If you have 20 minutes
Read the executive summary, then the three verdicts. That is the whole argument.
If you want to test the core premise
The thesis rests on frontier labs gating cyber models. Go straight to Access & gating — the evidence is more nuanced than the premise assumes, and it changes what you should build.
If you are deciding what to build first
If you are thinking about the model
If you are worried about the law
Germany, §202c and the Hackerparagraf is the one page a Berlin-based founder cannot skip. Then EU and US.
If you distrust any of this
Good. The verification ledger lists every claim that failed fact-checking, including several that other research passes asserted confidently and which turned out to be false.
Start here
The thesis, the verdicts, and how to read this.
Executive summary
The gating thesis is half right and closing. The defensible position is verification, not detection, sold into the EU compliance clock.
The three ideas, judged
Build/don't-build calls on the post-trained model, the PR review agent, and the open benchmark, each with falsifiable kill criteria.
Where the gaps actually are
Twelve unclaimed positions in the AI security market, scored on size, defensibility, and fit for a small European team.
The first 90 days
A week-by-week plan to produce one public artifact, three design partner conversations, and enough evidence to decide.
Landscape
The map of the space and what is actually driving demand.
The map
A taxonomy of the AI cybersecurity space: eight segments, who occupies each, and which ones are actually adjacent to each other.
What AI is actually doing to the threat landscape
A skeptical, tiered read of the demand side shows two real budget drivers and several headline stories that don't survive scrutiny.
Where value accrues
Seven layers from silicon to signed contract, which are commoditized, and the two where a small lab can hold ground.
Timeline 2023-2026
The chronological record of model releases, benchmarks, funding, and regulation that built this market, plus what's dated and coming next.
Players
Who exists, what they have raised, and what they can really do.
Offensive AI security companies
Autonomous pentest AI splits into four tiers competing on different bases, and the money is lopsided toward offense.
AI code security companies
Every layer of the PR-security stack has a vendor chasing it, and the frontier labs now compete directly with the startups.
AI SOC and detection companies
A crowded, capital-rich market selling alert-triage economics, sitting on the one benchmark that says reasoning may not help here.
The AI cyber lab category
Maps every organization already selling AI cyber-capability evaluation, how each makes money, and where a new entrant still fits.
What the frontier labs do themselves
OpenAI, Anthropic, Google DeepMind and Meta already occupy most of this market's layers, and they move in months, not years.
XBOW
XBOW is the best-funded autonomous pentest pure-play, with a genuinely qualified headline claim behind the hype.
Strix
Strix has the star count of a category leader and, in the one independent benchmark that exists, the results of a bottom performer.
Technology
Models, benchmarks, open source, data, and how to train.
Cyber benchmarks and evals
A master table of 20+ cyber benchmarks and a decisive verdict on whether building a new one is worth the founder's time.
The open-source stack
A map of every buildable open-source repo in AI security, which licenses will poison a commercial product, and what to actually assemble.
Open-weight security models
A survey of every open-weight cybersecurity model shows the market pays for classifiers, not chatbots.
Post-training playbook
A concrete plan, verifier design, and cost model for post-training a defensive cyber model in 2026.
Data, and whether a moat is possible
A full inventory of security training data shows vulnerability data is strong, malware and SOC data are years stale, and most published detection results are noise.
AIxCC: the closest thing to a proof
DARPA's two-year competition is the best public evidence of what autonomous find-and-fix can do, and where it still fails.
How these systems are actually built
The reference pipeline behind every working AI security agent, and why keeping execution deterministic is the one design choice that separates signal from noise.
Access & law
Frontier-lab gating, EU and US regulation, dual-use risk.
Is frontier-lab gating a real wedge?
The gating is real and wide today, but it is a relationship-and-timing moat rather than a technology moat, and it is closing.
EU regulation as a demand engine
A cluster of EU laws with hard dates turns vulnerability management into a mandatory, budgeted, recurring purchase.
Germany: §202c and the Berlin question
§202c StGB remains unreformed as of August 2026, which shapes corporate structure more than it should block the business.
The US picture
US demand signal is real but less reliable than the EU's right now — the flagship defense compliance driver was just suspended.
Dual-use risk and what it costs you
A single containment failure can end a small company even where it barely dents a frontier lab, so safety engineering is not optional.
Business
Buyers, pricing, gaps, economics, capital and talent.
Who buys, and what they pay
Engineering, not the CISO, signs most AI code-security deals, at $45-105/dev/month, and false positives are still the top complaint.
Unit economics and the compute bill
At $25/dev/month, a frontier-model deep review on every PR destroys margin; a well-optimized small model doesn't.
Who funds this and at what price
Maps every confirmed AI-security funding round and acquisition since 2024, who's writing the checks, and what 2026 investors actually reward.
Talent: the actual constraint
The hybrid security-research-plus-post-training hire is rarer and harder to compete for than capital in this category.
Go to market
The credible path is compliance-led EU enterprise pulled by CRA/NIS2, seeded by free open-source tooling and research-credibility inbound.
What could actually be defensible
Nine candidate moats tested against the two things that kill companies here: frontier labs shipping your feature, and commoditizing capability.
Reference
Sources, verification ledger, glossary, open questions.
Verification ledger
Every claim the adversarial fact-check pass could test, with verdicts, corrections, and what the fabrications would have cost us.
Open questions and the research backlog
A prioritised backlog of what desk research could not settle, with a cheap next step and a decision each answer would change.
Glossary
Precise definitions for the security, AI-safety, and regulatory terms used across this site, wikilinked to where each matters.
Sources and method
How this research was produced, what it drew on, where the source base is strong, and where it is thin.