Lasso Security has raised $30 million and launched LEAP, a CPU-based guardrail designed to screen routine AI traffic quickly while sending harder policy decisions to a separate self-hosted language model. The important story is not simply that another AI-security startup has raised capital; it is Lasso’s attempt to split security work into a cheap fast path and a slower judgment path.
Key takeaways
- Lasso Security says ClearSky led its new $30 million funding round, with Entrée Capital, iAngels, Singtel Innov8, Mindset and Swish Data participating.
- LEAP is a transformer-free guardrail that runs on ordinary CPUs; Lasso says it can handle most routine checks without calling a GPU-backed reasoning model.
- RAPID, a self-hosted language-model judge, is intended for the smaller set of requests that need nuanced interpretation of a plain-language policy.
- All latency, throughput, accuracy, cost, revenue, traffic and customer figures in the announcement are Lasso’s claims, not independently audited benchmarks.
The funding round was announced on September 2, 2026. Lasso did not disclose the round’s stage or the company’s valuation, so neither should be inferred. According to the company’s announcement distributed through GlobeNewswire, the money will support engineering, work with federal and regulated-sector customers, and go-to-market expansion in North America and Europe.
Lasso Security is an AI-security company founded in 2023. It develops controls for employees using generative AI, applications built on language models, and autonomous agents taking actions through tools. Its earlier $6 million seed announcement described protection across cloud and on-premise language-model environments. The new launch moves the pitch from discovering and monitoring AI use toward the cost of enforcing a decision on every prompt, response and agent action.
Lasso Security puts AI guardrails on a CPU fast path
Many AI guardrail systems use a second model to inspect an input or output. That approach can understand subtle context, but it also adds another inference request to the chain. At high volume, the guardrail can become a source of delay and compute expense even when almost every request is straightforward.
Lasso’s proposed answer is a two-engine design. LEAP handles the common, time-sensitive checks without a transformer model or a GPU. RAPID handles the exceptions that require a more interpretive decision. A routing layer chooses which engine receives a request. In principle, this avoids paying the highest inference cost for every interaction while retaining a language-model path for ambiguity.
This architecture matters because security controls must sit in the request path to stop harmful behavior before it reaches a model, database or external tool. A detector that is accurate but too expensive to apply broadly may be used only on samples. A detector that is fast but shallow may miss requests whose meaning depends on business context. Lasso Security is betting that a tiered system can reduce that tradeoff.
What LEAP and RAPID are designed to do
LEAP is not presented as a smaller general-purpose chatbot. Lasso describes it as a transformer-free security engine optimized for classification and enforcement on standard server processors. The company says it can inspect both what an AI system says and what an agent attempts to do. That distinction is increasingly important as agents move from answering questions to calling APIs, changing records or triggering workflows.
RAPID is the judgment layer. Lasso says it is a self-hosted language model used when a security decision requires interpretation of a natural-language policy. A rule such as “allow access to customer records only when the support case is open and the user belongs to the assigned team” may need more context than a simple keyword filter can provide. The platform’s router is supposed to reserve RAPID for those harder cases.
| Component | Compute approach | Intended role | Evidence caution |
|---|---|---|---|
| LEAP | Transformer-free, CPU-based | Routine inline checks at high volume | Performance figures come from Lasso |
| RAPID | Self-hosted language model | Nuanced plain-language policy decisions | Cost comparison comes from Lasso |
| Router | Tiered decision layer | Escalates only selected cases | Escalation rate was not disclosed |
The missing number is as important as the published ones: Lasso has not disclosed what share of real production traffic is escalated from LEAP to RAPID. That percentage would help buyers model actual compute savings. It would also show whether the fast engine handles the overwhelming majority of cases in practice or only in a curated benchmark.
Why CPU guardrails could change AI security economics
An AI application already pays for the model generating an answer or deciding an action. Placing another large model in front of every request can add inference cost, latency and infrastructure capacity. The problem becomes more visible with agents because one user instruction can produce a chain of tool calls, each of which may need a separate authorization decision.
That is the economic case for a CPU fast path. Standard CPUs are widely available in enterprise environments, including private clouds and air-gapped networks. If routine checks can run locally without reserving GPU capacity, an organization may be able to inspect more traffic. However, the cost argument only holds if the fast detector preserves useful accuracy across the attacks and policy violations that matter to the buyer.
Lasso says LEAP returns a decision in under five milliseconds and produces hundreds to thousands of times the throughput of existing guardrails. It also says the engine matches the detection quality of systems that require dedicated hardware. Those are company benchmark claims. Lasso has not published enough independent test data in the announcement to establish the workload, hardware, comparison set, false-positive rate or false-negative rate behind them.
The company also says RAPID can cost 100 to 200 times less than calling a commercial cloud language-model API. That figure is likewise a Lasso estimate and will vary with model choice, hardware utilization, prompt length, traffic shape and cloud pricing. Buyers should ask for a reproducible benchmark using their own policies and representative traffic rather than treating the headline multiple as universal.
AI agents make the enforcement problem harder
A chatbot can reveal sensitive information, but an agent may also modify a record, send a message, create code or initiate a transaction. This expands the security question from “Is this text safe?” to “Should this identity perform this action, on this system, with this data, at this moment?” The answer may depend on user permissions, task context and the sequence of earlier actions.
That is why Lasso Security’s new product belongs beside broader work on AI-agent supply-chain security, not only content moderation. Enterprises also need to consider prompt injection against connected business systems, tool provenance, identity controls and the permissions granted to each agent.
For Indian banks, insurers and financial platforms, the useful question is not whether a guardrail contains AI. It is whether the control can run inside the required environment, preserve audit evidence, respect data-residency choices and stop an action before execution. The Financial Stability Board’s concerns around AI concentration and cyber risk in finance underline why deployment architecture matters as much as model accuracy.
What the $30 million round does and does not prove
ClearSky led the round, according to Lasso. Entrée Capital, which led the company’s earlier seed financing, increased its position, while iAngels, Singtel Innov8, Mindset and Swish Data also participated. The company says it will add engineers and deepen work with federal and regulated-industry customers while expanding sales coverage.
Funding validates investor interest, not product performance. Lasso says its revenue grew more than 500% over the past 12 months, but it did not publish an absolute revenue figure. It says its platform protects tens of thousands of agents and billions of requests and actions each month, but those volumes were not independently audited in the announcement.
Lasso also names the US Department of Homeland Security, BMW, Leonardo Defense, Kaufman Rossin, eToro and Fiverr among representative users or customers. Those relationships are stated by the company. The announcement does not provide comparable deployment scope, contract values or implementation dates, so the names should not be treated as proof that every organization uses every Lasso module.
The most serious incident described in the launch material is also a company account. Lasso says its red team, during an engagement with an unnamed global healthcare provider, extracted patient records and could alter prescriptions. Because the provider, test conditions and remediation details were not disclosed, the episode illustrates the type of risk Lasso says it tests for; it should not be reported as an independently verified breach.
Questions buyers should ask before deploying Lasso Security
First, buyers should request precision and recall results for their own traffic, not only a single accuracy number. A guardrail that blocks legitimate prompts can disrupt work, while one that misses rare high-impact attacks creates false confidence. Testing should cover prompt injection, sensitive-data leakage, tool misuse and multi-step agent behavior.
Second, enterprises should measure end-to-end latency at the 95th and 99th percentiles under load. An under-five-millisecond engine claim does not necessarily equal an under-five-millisecond production decision after networking, policy retrieval, logging and escalation. They should also measure how frequently traffic moves to RAPID and what that does to total cost.
Third, the guardrail should complement rather than replace access controls. Managed authentication for connectors, such as the approach discussed in enterprise-managed MCP authentication, determines which identity can reach a tool. A runtime guardrail can inspect intent and content, but it should not become a substitute for least privilege, approval thresholds or audit logs.
The central test for Lasso Security is straightforward: can its CPU fast path safely make most routine decisions while its LLM path catches genuinely contextual cases without restoring the very latency and cost the architecture is meant to avoid? The $30 million round gives the company resources to answer that question in production, but only customer-specific, reproducible measurements can answer it for buyers.
Frequently asked questions
What is Lasso Security LEAP?
LEAP is a transformer-free AI-security guardrail that Lasso says runs on standard CPUs. It is designed to make fast decisions on prompts, outputs and agent actions without requiring a GPU-backed reasoning model for every request.
How much funding did Lasso Security raise?
Lasso Security announced a $30 million round led by ClearSky on September 2, 2026. Entrée Capital, iAngels, Singtel Innov8, Mindset and Swish Data participated. The company did not disclose the financing stage or valuation.
What is the difference between LEAP and RAPID?
LEAP is the CPU-based fast path for routine, high-volume security decisions. RAPID is a self-hosted language-model judge for requests that require more nuanced interpretation of plain-language policies. Lasso says a routing layer decides which engine handles each case.
Are Lasso’s speed and accuracy claims independently verified?
No independent benchmark was included with the announcement. The under-five-millisecond latency, throughput, accuracy and cost comparisons are Lasso’s claims. Prospective customers should test them with representative workloads, hardware, policies and attack cases.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
