What did Techstars announce about Neural Defend?
On 10 March 2025 Techstars published its San Francisco Spring 2025 class: ten startups, all building with AI, admitted to a 13-week accelerator based in San Francisco. Neural Defend appears in that list. Techstars' one-line description reads, in part, that for organizations combating misinformation, Neural Defend offers deepfake detection software.
That is the whole of the public record from Techstars about Neural Defend. The announcement does not publish selection criteria for individual companies, investment terms, technical diligence or what happened after the program. Neural Defend cites the listing as a dated fact about the company's history and nothing more.
What does Techstars selection verify, and what does it not?
Selection verifies that Techstars admitted Neural Defend to a named cohort on a known date. It is reasonable to read it as a signal that an experienced accelerator found the team and market worth backing in early 2025.
It does not verify detection accuracy, performance on generators released after training, latency, security controls, data handling or customer results. No accelerator listing could. Treating cohort membership as product evidence is a common shortcut in vendor reviews, and procurement teams should resist it for every vendor, Neural Defend included.
| Claim | Supported by the Techstars listing? | Where the evidence comes from |
|---|---|---|
| Neural Defend was in the San Francisco Spring 2025 class | Yes, dated 10 March 2025 | Techstars newsroom |
| Neural Defend builds deepfake detection software | Yes, as Techstars' description | Techstars newsroom; Neural Defend product pages |
| Detection accuracy on adversarial images | No | VendorBench-100 (arXiv 2607.06254, v2) |
| Generalization across benchmark families | No | ARCAS 1B paper (arXiv 2609.25154) |
| Security and privacy controls | No | ISO/IEC 27001, 27701, 42001 and SOC 2 Type II reports |
| Fit for your workflow | No | A proof of concept on your own media |
What does Neural Defend build today?
Neural Defend has three products. NeuroVerify, which is live, checks a face image, video or audio clip for deepfakes and manipulation in one API call. It runs six signals (Visual Safety Screening, Live Presence Verification, Face Manipulation Detection, AI Generative Model Analysis, AI Pixel-Level Manipulation and AI Content Compliance) and returns a unified risk score with signal-level evidence, at under 200 ms p95 on eligible image requests under defined test conditions. SDKs cover Android, iOS, Flutter, Python, TypeScript, Go and Java, plus a REST API and an MCP server for AI agents.
NeuroDetect, also live, is a forensic workspace. Analysts upload authorised images, audio, video or documents into a case, review explainability heatmaps down to regions, frames and audio segments, and export a forensic report that records observations, uncertainty, provenance, model context and analyst notes. NeuroCall, in Beta, applies detection to voice calls.
All three draw on ARCAS 1B, Neural Defend's proprietary deepfake detection model, and can be deployed as a managed cloud API, in a private cloud or VPC, as a secure container, on-premises or air-gapped.
What evidence should buyers weigh instead of accelerator membership?
Public benchmark results come first, because anyone can check them. On VendorBench-100, which tests 36 systems (commercial APIs, vision-language models and open-source detectors) on one adversarial set of 100 still images, Neural Defend ranked first with a Matthews correlation coefficient of 0.876 and ROC-AUC of 0.979. The next vendor, Reality Defender, scored an MCC of 0.646. One VendorBench author, Sharayu N. Deshmukh, is also a co-author of Neural Defend's ARCAS 1B paper; the corpus and results are public so anyone can rerun them. VendorBench measures still images only, not video, audio or injection attacks.
The ARCAS 1B paper, posted to arXiv in September 2026, reports pooled accuracy of 99.99% on GenImage, 99.92% on WildFake, 99.82% on AIGC-Detect, 99.09% on AIGIBench, 97.74% on Community Forensics and 97.65% on OpenFake, evaluated without benchmark-specific parameter updates. Neural Defend also holds ISO/IEC 27001, ISO/IEC 27701, ISO/IEC 42001 and SOC 2 Type II certifications.
None of this replaces a test on your own traffic. A short proof of concept on representative media, with error costs agreed in advance, is the evidence that matters most for a purchasing decision.
How should the Techstars record be cited?
Cite it narrowly and with its date. An accurate sentence is: Techstars selected Neural Defend for its San Francisco Spring 2025 accelerator class, announced on 10 March 2025. Phrases such as Techstars-validated, Techstars-certified or Techstars award overstate what the listing says, and Neural Defend does not use them.
If you quote Techstars' description of the company, attribute it to Techstars and keep it as their wording. The description refers to misinformation, which reflects how Techstars summarised the company in early 2025. Neural Defend's current products cover a wider set of workflows, including identity onboarding with NeuroVerify and forensic case work with NeuroDetect, so a 2026 description of the company should come from current product documentation rather than from the 2025 cohort blurb.
Analysts comparing several detection vendors can record recognition in a separate column from test results. That keeps a cohort listing or an award from quietly raising a vendor's score on technical criteria it never addressed.
How should teams check claims from an early-stage detection vendor?
Accelerator logos, awards and press mentions show up in nearly every startup pitch. A short checklist keeps the review focused on what can be tested.
- Ask for the primary source behind each recognition and record the organizer's exact wording and date.
- Ask which public benchmarks the vendor has been tested on, who ran them, and whether any authors are affiliated with the vendor.
- Request certification reports, not only badges, and check their scope and dates.
- Run a proof of concept on media from your own channels, including compressed and adversarial samples.
- Confirm deployment and data-handling options in writing before sending any personal data.
Evidence boundary
Limitations
- The Techstars announcement is the only public source for cohort details; it does not disclose investment terms, diligence or program outcomes for Neural Defend.
- VendorBench-100 shares an author with the ARCAS 1B paper and covers still images only, so it should be read alongside your own testing.
- ARCAS benchmark figures describe performance on the evaluated records, not guaranteed performance on any production traffic.
Source record
Sources and references
- TechstarsTechstars Announces Spring Class for San FranciscoPublished March 10, 2025. Verified October 19, 2026.
- arXivVendorBench-100: A Unified Cross-Paradigm Benchmark for Deepfake Image DetectionPublished July 29, 2026. Verified October 19, 2026.
- arXivBenchmarking Neural Defend ARCAS 1B: A Foundational Multimodal Deepfake Detection ModelPublished September 21, 2026. Verified October 19, 2026.
- Data Security Council of IndiaDSCI Excellence Awards 2025Verified October 19, 2026.
