Fundraising Fox

Salus

YC W26

San Francisco, US · Founded 2026 · 2 employees · Hiring · 10 known investors

Salus is a policy-aware proxy that intercepts and manages AI agent actions, allowing, clarifying, repairing, or escalating tool calls without requiring changes to the underlying agent or stack. It serves developers and teams building AI agents across LLMs and agentic frameworks.

Also known as usesalus.ai

Founders & leadership· Y Combinator alumni (W26)

Salus was founded in 2026 by Vedant Singh.

VSVedant Singh
Vedant Singhin𝕏FounderVedant Singh studied computer science at Stanford and now works on AI agent validation technology.

Investors · 10

Also in the syndicate · 3

İş Bank’s CVC armİş Bankası 100. Yıl Venture CapitalPitch Drive

Funding

SEC filings, press & company announcements

$4.7M disclosed across 1 round · 2023

Source: company announcements and press reports — follow each round's link for the claim.

Company profile

researched Aug 2026

Salus provides runtime control over the actions AI agents take. Rather than scoring outputs after the fact, the product sits at commit time: when an agent proposes an action, Salus checks it against policy and against the evidence gathered during the run before the customer's backend executes it. Decisions include allowing the action, asking for clarification, or blocking it, and each intervention produces a decision receipt that records the proposed action, the governing policy, the missing evidence and the recommended safe retry. When an action is blocked, the agent receives structured feedback intended to guide a corrected attempt.

The company describes the product as an API that wraps around an existing agent, installable via a pip install and a few lines of code. Validation rests on two mechanisms: an evidence cache per run, holding prior tool-call outputs and conversation history, against which proposed actions must be grounded; and policies authored in YAML, markdown or plain English that are compiled into runtime checks that can be version-controlled, diffed and shadow-tested. Additional features include PII detection, budget and loop protection, idempotency, human-in-the-loop escalation and content moderation. Salus also bundles eval generation and observability in the same product, generating adversarial and realistic evals with context of the customer's agent domain.

Salus is a Y Combinator Winter 2026 company based in San Francisco, categorized under artificial intelligence, developer tools, B2B, API and infrastructure. Team size is listed as two, with an open founding GTM/BDR role in San Francisco or remote.

Founding story

Salus was founded in 2026 and joined Y Combinator's Winter 2026 batch. Its launch post describes two founders, Kevin and Vedant, former roommates at Stanford who studied computer science and are now working on Salus full-time. Vedant Singh is listed as Founder/CTO and described as an AI researcher formerly at Stanford CS. The founding motivation described is the gap between reactive tooling — evals, output scoring and observability — and the absence of a mechanism that inspects an agent action at the moment it is about to execute, illustrated by examples such as an agent issuing a refund without looking up the order or emailing a lead with hallucinated data, with the error discovered only hours later.

Business model

Salus sells software infrastructure to companies deploying AI agents. Customers integrate the API into their agent stack, typically starting with a single protected action route, and Salus makes the runtime allow/clarify/block decision before the customer's backend executes the action. Onboarding is described as a pip install plus a few lines of code, and prospective customers are engaged through demos and a pilot application. Pricing terms are not stated.

Traction

Salus reports internal controlled benchmark results reproduced across three matched repetitions. On AgentDojo banking with gpt-4o-mini, mean prompt-injection attack success fell from 47.7% to 0.69%, a 98.6% reduction, with attack success held to 1.4% or less in every repetition and 0% in the best single run. On AgentDojo workspace with gpt-4o-mini, utility rose from 37.5% to 64.4% while attack success fell from 17.1% to 1.73%. The company also cites tests on tau-squared-bench, where agents with Salus are said to follow policies more reliably at up to 60% lower cost, and ODCV-Bench, where it reports a 52% average reduction in misalignment across 12 frontier models. In its own benchmarks, 58% of blocked actions are reported to recover and complete the task correctly. Customer counts and revenue are not disclosed; the company is soliciting pilots.

Latest developments

Salus participated in Y Combinator's Winter 2026 batch, with Ankit Gupta as primary partner, and published a launch post describing the runtime guardrails product, its benchmark results and immediate onboarding for teams deploying agents. The company is running a pilot program starting in shadow mode on a single protected action route and is recruiting its first go-to-market hire.

Full profile — market position, technology, go-to-market, geography

Market position

Salus positions itself against evals, output scoring and observability tooling, which it characterizes as reactive measures that reduce the likelihood of agent errors but do not inspect an action as it is about to execute. Its stated claim is that runtime action validation is the missing layer, and it combines runtime validation, eval generation and observability in a single product.

The company's stated points of difference are intervention at commit time rather than after execution; requiring proposed actions to be grounded in an evidence cache of prior tool outputs and conversation history; policies authored in plain English, markdown or YAML that compile into version-controlled, diffable, shadow-testable runtime checks; structured feedback that lets blocked actions self-repair; and an auditable decision receipt for every intervention. Integration requires no changes to the underlying models or orchestration frameworks.

Technology

The system intercepts proposed agent actions at commit time and evaluates them against two inputs: an evidence cache maintained per run, containing all prior tool-call outputs and the full conversation history, and explicit policies written in YAML, markdown or plain English that Salus compiles into runtime checks that are version-controlled, diffable and shadow-testable. Outcomes are structured decisions (for example ALLOW or CLARIFY) accompanied by a receipt naming the policy, the missing evidence and a specific correction; blocked actions return structured feedback to drive a self-repairing retry. Supplementary controls include PII detection, budget and loop protection, idempotency, content moderation and human-in-the-loop escalation. The product is model- and framework-agnostic, listing compatibility with OpenAI, Anthropic and Gemini and any tool-calling LLM; LangChain, LangGraph, CrewAI, LlamaIndex or a custom loop; voice stacks including Retell and Vapi; and tools exposed via MCP servers, REST, SDKs or internal services.

Go-to-market

Salus markets directly to engineering teams deploying agents, offering a demo booking and a pilot application on its website and onboarding via a lightweight SDK install. Deployment follows a staged path intended for one high-risk action route: replaying recent traces to inspect proposed actions without customer impact, authoring a policy for that action, running Salus in shadow mode alongside live behavior for comparison, then promoting the route to live enforcement. The company asks prospects to bring a single workflow touching money, customers or systems of record. It is hiring a founding GTM/BDR in San Francisco or remote at $35-$45 per hour, open to new graduates.

Developers and teams deploying AI agents that take consequential actions — issuing refunds, writing records, booking appointments, contacting customers — including builders of in-house agent loops and voice agents. Y Combinator categorizes the company as B2B, developer tools, API and infrastructure.

Geography

Headquartered in San Francisco, California, with a remote-eligible open role.

Compiled by commissioned research from 8 cited public sources — announcements, filings, and press listed under research sources below.

Key figures

latest reported
Attack success rate ceiling across repetitions (AgentDojo banking)Jan 20261.4%
Average misalignment reduction on ODCV-Bench across 12 frontier modelsJan 202652%
Blocked actions that recover and complete the task correctly (internal benchmarkJan 202658%
Cost reduction on tau-squared-bench with SalusJan 202660%
HeadcountAug 20264
Reduction in prompt-injection attack success rate (AgentDojo banking, gpt-4o-minJan 202698.6%
Team sizeJan 20262 employees
Utility improvement (AgentDojo workspace, gpt-4o-mini, 3 repetitions)Jan 202627%

Company-reported or press-reported figures, each dated to when it was claimed — not independently audited.

Competitors · 2

by search overlap

Companies competing with Salus for the same Google search keywords, organic and paid, via search-intersection analysis.

Timeline · 2

launches, deals, and filings
Jan 2026
Selected for Y Combinator Winter 2026 batch

Salus is listed as an active Y Combinator Winter 2026 company based in San Francisco, with Ankit Gupta as primary partner.

source ↗

Jan 2026
Y Combinator launch: runtime guardrails that validate AI agent actions before they execute

Company launch post introducing Salus as an API that wraps an agent and checks its actions at runtime, blocking incorrect ones and returning structured feedback to guide retries, with evidence-cache grounding, policy compilation, PII detection, budget and loop protection, idempotency, human-in-the-loop escalation and content moderation, plus bundled eval generation and observability.

source ↗

Dated company events from announcements, filings, and press; legal rows summarize public dockets and regulator releases.

In the news

Research sources · 8

primary sources listed

8 public sources were cited for this profile; the first-party ones are listed here.

Frequently asked questions

What does Salus do?
Salus is an API that validates AI agent actions at runtime, blocking unsafe tool calls and feeding back corrections.
Who founded Salus?
Salus was founded by Vedant Singh in 2026.
Who are Salus's investors?
Salus's investors include Y Combinator, 500 Emerging Europe, Anterra Capital, Collective Spark, Northzone, Pitchdrive, Wamda Capital.
How much funding has Salus raised?
Salus has disclosed $4.7M raised across 1 round.
Where is Salus headquartered?
Salus is headquartered in San Francisco, US.