Meeting the new pressures of AI security

AI applications don’t stand still. Inputs drift week by week, usage goes global overnight, and end users expect every interaction to remain sub-second fast. Security teams are caught in the middle: block too much, and you frustrate users; miss too much, and you take on risk.

But there’s a quieter and costly security problem: noise. Every AppSec team has wrestled with tools that drown analysts in irrelevant alerts, creating fatigue and skepticism. That history matters. When detection quality slips, even great security loses trust.

With our Fall’25: Adaptive at Scale release, Lakera Guard tackles these pressures head-on. This release introduces three capabilities—adaptive calibration, multilingual moderation, and consistent latency—that work together to keep security smart, global, and fast, while advancing upon the Lakera precision and confidence security teams expect.

Adaptive to your data

Static rules and generic threat models quickly lose relevance as agents evolve. That’s why we’ve introduced adaptive calibration. Instead of treating all traffic the same, Guard now learns what “normal” looks like for your applications. It automatically adjusts flagging criteria to reduce noise and surface the truly suspicious.

It is well known that in Application Security, false positives aren’t just a nuisance, they’re a credibility killer. Every practitioner has felt the fatigue of noisy tools flooding their queue with “findings” that turn out to be irrelevant or wrong. Early SAST tools built that reputation, and we haven’t forgotten. That’s why our new per-application model tuning is such a breakthrough. By tailoring detection models to customer-specific data, we’re achieving false-positive rates as low as 0.1-0.2%, a level that redefines accuracy for modern AppSec.

For context, traditional or open-source SAST tools often hover around 20%, and even highly tuned commercial systems rarely drop below 1-3%. Cloud security alerts typically fall between 1-5%. Reducing that to near-zero isn’t just a technical improvement—it means less wasted analyst time, fewer SIEM floods, and more trust in every alert that does appear.

Early results show noise dropping up to 90% within the first three weeks of use. That means fewer interruptions, less wasted analyst time, and faster mean-time-to-signal when something’s actually wrong.

Global by default

AI threats are not limited to English. Offensive content and abuse arrive in dozens of languages, often leveraging slang or cultural nuance that defeats naive filters. Our new multilingual content moderation model brings accurate detection across 100+ languages - the same comprehensive coverage that already powers Lakera Guard’s prompt defense.

While prompt defense protects the model from malicious inputs, content moderation protects users from harmful outputs — both now powered by the same multilingual foundation.

The benefit is straightforward: one policy layer can now protect your global user base, without requiring you to stitch together per-locale rules. Customers already testing the model are seeing materially fewer moderation errors and escalations across top languages.

We benchmarked our new moderation model on Lakera’s internal Challenging Moderation Benchmark, a multilingual dataset designed to expose edge cases that often trip up moderation systems. Despite the difficulty, the new model achieved 60% fewer false positives overall while improving accuracy by over 200% in low-resource languages like Hindi and Chinese.

Speed at scale

Strong security is worthless if it adds friction. Users expect every agent interaction to complete in real time, which makes tail latency as important as average latency. With our new consistent latency architecture, Guard now runs detectors in parallel and prioritizes intelligently, so you can keep protections on without slowing down your app.

Lakera Guard keeps latency tight and predictable across every workload: sub-40ms for short prompts, and under ~200ms for the most thorough, book length requests.

Availability and rollout

  • Adaptive calibration: Rolling out on a per-customer basis for SaaS customers
  • Multilingual moderation: GA for Tier-1 languages; Beta for long-tail coverage. Available for both SaaS and self-hosting customers.
  • Consistent latency: Live for all SaaS customers. On the roadmap for self-hosting customers.

Transparency in testing

We believe benchmarks only matter if you can trust the methodology. For this release we measured against representative workloads across industries, fixed the test harness to ensure comparability, and reported both averages and tail latencies.

Looking ahead

Adaptive calibration, multilingual moderation, and consistent latency are more than features—they’re steps toward our larger mission: protecting every AI agent you build and run. By making Guard adaptive, global, and fast, we’re ensuring that security keeps up with the scale of modern AI.

-db1-Ready to see it in action? Book a 15-minute demo or explore the new capabilities directly in your Guard console.-db1-