Stephen Ward does not talk about exits the way a fund manager tells a success story. "You have to sell on the way up," says the co-founder of Brightmind Partners. In January, CrowdStrike agreed to buy the identity security firm SGNL for $740 million — 11 months after Brightmind led its Series A. For a two-year-old fund with a $250 million debut, that is the proof point every limited partner wants to see.

Ward is now raising a second fund of roughly the same size, this time without his co-founder. Gur Talpaz left in April, along with the firm's third employee, Tayler Sipperly. In August the two co-founded Cognition with Elia Zaitsev, CrowdStrike's longtime CTO; it is reportedly targeting a $170 million first fund. One partnership became two, and one of them is a competitor.

The interesting part of Ward's story is not his résumé, unusual as it is — Coast Guard, nine years in the Secret Service, CISO roles at TIAA and Home Depot, then venture. It is the numbers he brings with him. According to PitchBook, early-stage cybersecurity startups have out-raised their later-stage rivals this year for the first time in more than a decade: $4.7 billion from seed through Series B by the end of September, already more than all of 2025. The number of deals fell from 358 to 248. The average seed round rose from $5.54 million to $8.4 million.

Fewer deals, more money per deal. That is the fingerprint of a market where investors would rather place many early bets on AI themes than pay up for later-stage valuations. Ward himself says he is happy to sell early rather than wait for an IPO — a rational stance in a segment crowded with strategic buyers like Palo Alto Networks, CrowdStrike and Zscaler.

His sharpest finding, though, concerns a label. PitchBook counted 278 deals in the first half of 2026 bucketed as AI security. Ward's team reviewed them by hand and found about 30 that were actually doing AI security. The rest, he says, was ordinary security software with a new phrase in the pitch deck, or companies repackaging conventional products around a model.

That is not a footnote. If a third of a segment's growth is renaming, then "AI security" is useless as a market measure, and LPs allocating against it may be buying something other than what they think. Ward instead favours models that run locally on customer systems, because real-time protection — he calls it "zero second" — cannot afford a round trip to the cloud, and he names Qualys and Tenable, 27 and 24 years old respectively, as incumbents he considers vulnerable.

How Brightmind works is also notable. Its CTO built the AI agents that now source most of the firm's deal flow — a deliberate refusal to hire an analyst bench, which keeps management fees down. Whether that improves the hit rate or only the volume will show in the next 14 investments.

For founders, the story carries two lessons: there is plenty of early capital for security, but less willingness to pay a premium for it. And a label is not a category.