Latest / UpNext AI / Ballistic Ventures’ Cybersecurity Funds, Cisco’s OpenAI Defense, and Parkinson’s Screening
Transcript
- Welcome to the UpNext AI podcast. It's Wednesday, November 4th, 2026, and here's what matters in AI today. TechCrunch reports that Ted Schlein’s cybersecurity investment firm, Ballistic Ventures, is seeking to raise $490 million across two new funds. Regulatory filings show a target of $360 million for its next flagship fund and $130 million for a separate vehicle. Neither fund had recorded any sales when the filings were made, so these are fundraising plans, not completed raises. The plans come as AI changes both sides of the security market. Attackers can use it to work faster and at greater scale, while companies are building tools to monitor and test increasingly autonomous software. For a measure of investor interest, JPMorgan Chase, citing PitchBook data, says AI-enabled companies accounted for about 72 percent of U.S. cybersecurity venture deals through May of this year, compared with 36 percent in 2019. There is a concrete buyer story alongside the fundraising: TechCrunch also reports that CrowdStrike plans to acquire Fiddler AI, which makes agent-observability software, and Dreadnode, which helps automate AI security testing. Ballistic previously backed Pangea, another company bought by CrowdStrike. Together, the fund filings and acquisition plans show investors and security vendors positioning around new demands created by AI agents. Ballistic declined to comment on its fundraising plans. That shift is also changing how established security companies work. In a company account published by OpenAI, Cisco describes combining its network telemetry with OpenAI models to help security teams connect signals across their systems. Cisco’s Jeetu Patel says a pattern can become meaningful when signals from different parts of a network are considered together. Cisco’s existing controls can restrict which services communicate and isolate affected systems while other services keep running. The company sees model-assisted investigation as a way to help teams identify vulnerabilities and respond faster within their existing scanning workflows. Those improvements are Cisco’s account of its work, rather than an independently measured speed comparison. Agents add another job for defenders: distinguishing malicious interference from an agent carrying out a legitimate goal in an unintended way. Patel says teams need security telemetry alongside information about agent behavior to make that distinction. Through the Daybreak Defense Network, Cisco and OpenAI are exploring faster exposure triage and controls in Cisco Cloud Control. The useful distinction for operators is that watching what an agent does belongs alongside watching what reaches it. A quick disclosure: we use models from OpenAI to help write the UpNext AI podcast. A different use of language models appears in Nature Aging. A paper describes screening for Parkinson’s disease using answers to self-reported questionnaires, rather than requiring the person to supply a voice recording or another sensor measurement. The researchers make the code for their Quest2Dx approach public, including model definitions and training and evaluation scripts. Questionnaire-based screening could give researchers a way to examine information people can report themselves. It also calls for a clear distinction between a screening result and a diagnosis. Anyone assessing this approach for a health workflow would need its measured performance and the population in which it was evaluated before deciding what role it could play. For teams comparing language models, even running the tests can be inconsistent. Benchmark authors often use different evaluation conventions and custom scripts, making results harder to reproduce. A paper published yesterday proposes DEP, a decentralized evaluation protocol and toolkit, to give those tests a shared process. The design puts benchmark files, scoring logic and answers on a server, which can run locally or remotely. In a remote setup, users do not receive the ground-truth answers; that arrangement is intended to reduce benchmark leakage. The toolkit also supports resuming interrupted runs and handling concurrent requests. The researchers evaluated multiple language models across benchmarks and report that DEP reduced the cost of deploying benchmark evaluations. They say they had adapted more than 60 benchmarks by February of this year. The result is about making evaluations easier to deploy and keep consistent; the paper’s abstract does not quantify the cost reduction. ...Are you building apps with voice? Elevate your app's voice capabilities with ElevenLabs. Their API is a game changer for embedding dynamic, responsive voice interactions in your applications, providing unprecedented realism, flexibility and latency. Visit up next dot fm slash eleven to check out their latest offerings. ... A few more developments before we go. TechCrunch reports that Anthropic is working on FlashProve, an open-source system intended to cryptographically verify which model produced an output. Researchers have released a technical specification and working code, but say the work is ongoing and there are no plans to deploy it at a lab. For schools, Google says it is giving administrators more control over its AI offerings. The Information reports a consequential condition: schools that want Google’s search service must use an AI-enhanced version. Search and AI controls therefore come as a linked decision for those schools. Anthropic has released Message Threads for its Messages API. Agents can send changes to a conversation instead of retransmitting its entire history at every turn. The Decoder reports a 35 percent reduction in time to first token for long conversations—that measures when a response starts, not when it finishes. And The Information reports that OpenAI has avoided publicly taking a position on AI chip sales to China out of concern about alienating Nvidia. Its account points to Nvidia’s control over the chips developers need as a source of leverage in that policy debate. Before we wrap up, a quick note: this podcast is generated with the assistance of AI and is intended for informational purposes only. All referenced articles, research, and commentary remain the property of their original authors and publishers. And that's your briefing for today. If you enjoyed this episode, share it with a friend or colleague—and follow UpNext AI on Apple Podcasts, Spotify, or wherever you get your podcasts! Full source links are in the episode notes, and we'll be back tomorrow with what's up next!