Silmaril
Two-person YC startup selling a runtime layer that blocks prompt injection on AI agents.
On the record
Checkpoint: First independent red-team benchmark of Silmaril's detection claims, or a fifth security acquisition in this space
The call
Silmaril sells a thin, easy-to-swap security layer into a market where bigger security and cloud companies already ship the same feature for free.
Silmaril builds software that protects AI systems from attacks and misuse by catching bad prompts before they cause harm. The problem is that larger companies like Check Point, Palo Alto Networks, and Cisco have already bought their way into this space with bigger acquisitions, while Amazon, Microsoft, and Google are giving away similar protection for free inside their own cloud services. What could change everything is if independent security researchers publish real test results showing Silmaril actually catches attacks that competitors miss, or if the whole AI safety market suddenly shrinks because fewer companies are actually building AI agents.
Why
01
AI pressure: Cloud platforms and frontier labs (AWS, Azure, GCP, OpenAI) shipping native guardrails. AWS, Azure, GCP and OpenAI now ship built-in prompt-injection filters for free, cutting the case for a paid layer. Could AI do the core job? Yes, likely.
02
Bigger players: Check Point (Lakera), Palo Alto Networks (Protect AI), Cisco (Robust Intelligence), OpenAI (Promptfoo). Four security and AI giants already bought or built runtime AI-security products before Silmaril left pre-seed. Could a giant copy it? Yes, easily.
03
What protects it. Little beyond YC backing and a self-improving classifier loop; no independent test confirms it beats free rivals.
What would change this call
A fifth AI-security acquisition (after Lakera, Protect AI, Robust Intelligence, Promptfoo) by Dec 2026. Would show incumbents are closing the market before Silmaril can scale.
Independent, non-vendor red-team benchmark comparing Silmaril to Lakera Guard or Protect AI. Would confirm or debunk Silmaril's self-reported 95.6% vs 74.3% accuracy claim.
AWS, Azure, or GCP ships a native guardrail feature matching Silmaril's self-healing claim. Would show cloud platforms absorbing Silmaril's core pitch for free.
Independent, non-vendor red-team benchmark comparing Silmaril to Lakera Guard or Protect AI. Would confirm or debunk Silmaril's self-reported 95.6% vs 74.3% accuracy claim.
AWS, Azure, or GCP ships a native guardrail feature matching Silmaril's self-healing claim. Would show cloud platforms absorbing Silmaril's core pitch for free.
Key numbers
- Two-person founding team (CEO, CTO)
- Vendor claims 96% attack block rate vs 61% rivals
- Vendor claims $28M in damages prevented for customers
Named risks
- Category actively being rolled up by Check Point, Palo Alto Networks, Cisco, and OpenAI acquisitions.
- Cloud platforms ship equivalent guardrails natively and free, eliminating need for point solution.
- All performance and traction claims are self-authored with no independent audit as of June 2026.
- Two-person team with only $500K raised versus well-funded, larger-headcount rivals.
- Product positioning shifted within two months (April to June 2026), suggesting unclear market fit.
Rivals, as read in July 2026
| Rival | Direction | Evidence |
|---|---|---|
| Lakera (Check Point) | losing to | Check Point acquired Lakera for ~$300M Sept 2025; Gandalf dataset has 80M+ verified prompts vs Silmaril's smaller dataset |
| Protect AI (Palo Alto Networks) | losing to | Acquired by Palo Alto Networks in 2025, gaining enterprise distribution Silmaril lacks |
| Robust Intelligence (Cisco) | losing to | Acquired by Cisco in 2024, folding agentic-AI security into a networking giant |
| Promptfoo (OpenAI) | losing to | OpenAI acquired Promptfoo March 2026, pulling red-teaming tooling into a frontier lab |
| AWS/Azure/GCP native guardrails | losing to | Bedrock Guardrails, Azure AI Content Safety, Model Armor ship free native scanning as of Dec 2025 |
| Singulr AI | contesting | ~47-employee Palo Alto rival actively promoting at RSA Conference April 2026 vs Silmaril's 2-person team |
Questions people ask
- Is Silmaril profitable in 2026?
- No public revenue or profitability figures exist. Silmaril has raised only $500K via convertible note as of a June 17, 2026 CB Insights snapshot, and its only traction claim is a vendor-reported $28M in prevented damages, unaudited by any third party.
- Who founded Silmaril?
- Silmaril was founded by CEO Aum Upadhyay and CTO Eduardo Velasco, a two-person team, according to Silmaril's Y Combinator company page published June 18, 2026. The company went through YC's Spring 2026 batch, presenting at Demo Day around June 16-18, 2026.
- Could AI replace Silmaril's product?
- Yes, largely. Silmaril's own product is an AI classifier, and cloud platforms AWS Bedrock Guardrails, Azure AI Content Safety, and Google Cloud Model Armor ship equivalent prompt-injection scanning natively and for free, per a December 31, 2025 field guide, directly threatening point solutions like Silmaril.
- Who is Silmaril's biggest competitor?
- Lakera, whose Gandalf dataset holds 80M+ third-party-confirmed adversarial prompts, was acquired by Check Point Software in September 2025 for roughly $300M, giving it enterprise distribution and a data moat larger than Silmaril's claimed dataset, per CB Insights, June 28, 2026.
- How much funding has Silmaril raised?
- Silmaril has raised $500K total via a convertible note, with Y Combinator listed as an investor, per a CB Insights snapshot dated June 17, 2026. No priced round, institutional lead, or valuation has been publicly disclosed as of that date.
- Are Silmaril's attack-blocking benchmarks verified?
- No. Silmaril claims 96% attack blocking versus 61% for 'leading guardrails' (April 13, 2026 YC Launch) and separately 95.6% accuracy versus Lakera Guard's 74.3% (growwithsilmaril.com, June 18, 2026), but both figures are self-authored with no independent third-party audit found.
Timeline
- 2026-06-30GitHub SDK silmaril-security-sdk releases v0.4.1-0.4.2, showing active development
- 2026-06-18YC Spring 2026 Demo Day; product repositioned as runtime security for self-improving AI
- 2026-06-17CB Insights snapshot shows $500K total raised via convertible note, status Alive
- 2026-04-13YC Launch as self-healing prompt injection defense; claims 96% vs 61% blocking
- 2026-03-01OpenAI acquires Promptfoo, absorbing red-teaming tooling into a frontier lab
- 2025-09-01Check Point acquires Lakera for roughly $300M, consolidating prompt-injection defense
- 2025-01-01Palo Alto Networks acquires Protect AI during 2025 (exact date unspecified)
- 2024-12-31Cisco acquires Robust Intelligence, an ML/agentic security vendor
Revision history
27 July 2026
First assessed. Losing its protection. The text above is fixed as of this date and is not edited afterwards.
In the same band
Sources
- Products, Competitors, Financials, Employees, Headquarters Locations (2026-06-17)
- The 11 standout startups from YC's Demo Day, according to VCs (2026-06-18)
- Silmaril: Runtime security for AI that self-improves (2026-06-18)
- Launch YC: Silmaril: Self-healing Prompt Injection Defense (2026-04-13)
- Runtime Security for AI (2026-06-18)
- Silmaril-Security/sdk-python (2026-06-30)
- Products, Competitors, Financials, Employees, Headquarters Locations (2026-06-28)
- Stop Letting Models Grade Their Own Homework (2026-01-18)
- Best AI Security Tools 2026 (2026-07-20)
- What Is Prompt Injection? 2026 Defense Field Guide (2025-12-31)
- Innovator Spotlight: Singulr AI (2026-04-22)
- The rise of the agent runtime (2026-07-21)