Vijil

Vijil sells three linked products. Evaluate is a QA/red-teaming engine: point it at an agent (via API, a local executor tunneled through ngrok, or a hosted endpoint) and it runs it against a library the vendor says exceeds 200,000 adversarial and benchmark prompts, then returns a 'Trust Score' broken into reliability, security, and safety sub-scores with per-failure explanations. Dome is a runtime guardrail layer — a Python library or Docker sidecar that filters inputs/outputs against natural-language policies you define, with a claimed sub-300ms latency add. Darwin, the newest piece, is pitched as reinforcement learning that adapts an agent's defenses from production telemetry and evaluation failures over time. A fourth capability, Discover, scans code repos and cloud infrastructure to inventory an org's agents, including unsanctioned ones.

An ngrok engineer built a small git-log-summarizing agent and ran it through Vijil Evaluate. The tool caught a real vulnerability: the agent ignored injected 'noisy' user context and leaked confidential data when prompted to, dropping its passing score from 97.56 to 95.51 after the developer actually wired the injected context into the prompt (the first, higher score reflected the agent silently discarding Vijil's test inputs, which is not the same as passing them). That's a specific, reproducible account of the tool surfacing a real failure mode rather than rubber-stamping a pass — a useful signal for reliability-conscious buyers.

A 2025 Mindgard research paper testing prompt-injection and jailbreak evasion techniques (emoji smuggling, zero-width characters, homoglyphs, adversarial ML perturbations) against six commercial and open-source LLM guardrails — Microsoft Azure Prompt Shield, Meta Prompt Guard, Nvidia NeMo Guard, Protect AI v1/v2, and Vijil's own prompt-injection guardrail — found every one of them bypassable to some degree, with some attacks achieving 100% evasion against multiple vendors including Protect AI v2 and Azure Prompt Shield. Mindgard's overall finding was 'there is no best guardrail'; Vijil was one of six tested, not singled out as worst or best. Vijil's marketing claims Dome is 'up to 2x more accurate than Amazon Bedrock guardrails' — a vendor claim we could not independently verify.

Vijil Evaluate publishes three tiers on its site. Individual/usage-based is free and includes 1,000 credits, 25+ benchmarks, Trust Score, the Garak red-teaming engine, a playground, and support for custom harnesses and agents, plus email/Slack support. Team is usage-based ('pay-per-eval', no dollar figure published) and adds shared harnesses, evaluations, billing, and API keys. Enterprise is an annual subscription, contact-sales only, adding on-prem deployment, MCP/A2A testing, SSO/RBAC, and dedicated 8x5 support. Academic and research use is free forever by application. A separate FAQ on the same page says the free trial runs '3 months,' which sits oddly next to the pricing table describing an always-free Individual tier with credits — worth clarifying with sales before assuming either applies to your use case. Dome, the runtime guardrail product, is offered free for 30 days on request but has no published tier ladder beyond that.

Founded in 2023 by Vin Sharma (CEO, ex-GM/Director of Engineering at Amazon SageMaker) and Zdravko Pantic (Head of Engineering, ex-AWS AI, previously led PyTorch/TensorFlow/SageMaker Training teams). The company raised $17M in a round led by BrightMind Partners with participation from Mayfield and Gradient, announced November 25, 2025, bringing total raised to $23M. It was named to Gartner's 2025 Cool Vendors in Agentic AI Trust, Risk and Security Management (corroborated independently by SiliconANGLE citing the Gartner document) and claims a CB Insights 'Most Innovative AI Startups 2025' listing, which we could not independently verify beyond Vijil's own site. Advisors include Leon Derczynski (creator of the Garak red-teaming tool and NeMo Guardrails, and OWASP LLM Top 10 lead) and Ruslan Salakhutdinov (CMU professor, deep-learning researcher) — a notable technical bench for an evaluation-focused security startup.

DigitalOcean published a detailed customer story describing Vijil working with its customer Autonoma to take an IoT troubleshooting agent from a hallucinating prototype to production in one week — including specifics like switching the base model from Mistral Nemo Instruct to Llama 3.1 8B to reduce prompt-injection susceptibility, and splitting English/German knowledge bases to fix language-switching failures. Vijil's own site names SmartRecruiters, DuploCloud, and DigitalOcean-platform developers as production users; we corroborated the DigitalOcean/Autonoma relationship independently but did not verify the SmartRecruiters or DuploCloud claims beyond Vijil's own marketing.

No independent review sites (G2, Capterra, TrustRadius) turned up coverage at this company's stage — normal for a two-year-old startup, but it means a buyer has no third-party satisfaction signal beyond the single DigitalOcean case study and the ngrok hands-on account. Vijil publishes no public changelog or status page, so there's no way to check release cadence or incident history from outside the product. The Darwin (continuous-learning) and Discover (agent-inventory) products are described only in marketing copy with no independent usage accounts found.