Case studies / Guardrails

Guardrails

Stopping hallucinations, unsafe answers and prompt injection; moderation.

Approach

21 write-ups

Vercel4 Nov 2026NewGuardrails
Vercel token scanning expands to more public sources
Problem Publicly exposed Vercel tokens let attackers access projects and incur unexpected costs.
Solution They created an automated scanner, partnered with Truffle Security and added a monthly web crawl check, to detect and revoke leaked tokens across many public sources.
Hundreds of tokens revoked
Read on vercel.comDetails
Vercel3 Nov 2026NewGuardrails
Vercel Security Review now in public beta
Problem Developers needed automatic detection of code vulnerabilities on every push.
Solution Created an AI‑driven Security Review that scans code after each GitHub push, flags issues, rates severity, and can generate fix pull requests via the Vercel Agent.
Read on vercel.comDetails
Replit28 Oct 2026Guardrails
Replit’s Architecture Behind Secure Agent Work
Problem Agents could be exploited to cause damage, so companies need secure guardrails and monitoring.
Solution Replit added isolation, firewall rules, and policy‑driven controls in its Agent Habitat platform to limit what agents can read or write and to automatically enforce security rules.
Read on replit.comDetails
Vercel23 Oct 2026Guardrails
How Vercel Agent accesses production safely
Problem Agents need to act on production without over‑granting permissions, which could cause accidental damage.
Solution Vercel Agent creates a step‑by‑step plan that shows required actions and permissions; users approve each plan before any changes are made.
Read on vercel.comDetails
NVIDIA21 Oct 2026Guardrails
Inside NVIDIA Sentry: Secure AI Agents at Runtime
Problem AI agents could bypass security controls and misuse tools or data.
Solution They created NVIDIA Sentry, a hardware‑based framework on BlueField‑4 DPUs that watches agent actions, checks requests, and enforces policies outside the agent’s reach.
Read on developer.nvidia.comDetails
Salesforce30 Sep 2026Guardrails
How Deterministic Controls Turn AI Output Into Reliable Prompt Templates
Problem Open-ended AI instructions produced inconsistent and unsafe outputs, making prompt templates unreliable.
Solution Implemented deterministic controls that turn free-form instructions into stable, safe prompt templates for the Prompt Builder.
Read on engineering.salesforce.comDetails
NVIDIA28 Sep 2026Guardrails
NVIDIA Open Agent Safety Platform: A Reference for Continuous In-Silicon Agent Monitoring
Problem AI agents were breaking out of their test environments and accessing unauthorized systems.
Solution Created OpenShell, an open‑source runtime that runs agents in a sandbox with kernel‑level isolation and continuous monitoring.
Read on developer.nvidia.comDetails
NVIDIA28 Sep 2026Guardrails
Add Runtime Controls to AI Agents with NVIDIA OpenShell
Problem Agents could access too many resources, risking data leaks or unintended actions.
Solution They created OpenShell, a sandbox that enforces policies, limits file and network access, and protects credentials for any AI agent without changing its code.
Read on developer.nvidia.comDetails
NVIDIA22 Sep 2026Guardrails
Enabling Private High-Performance Production AI Inference with NVIDIA Confidential Computing
Problem Customers need to run LLM inference on sensitive data without exposing it.
Solution They created confidential virtual machines and encrypted GPU links, then adapted TensorRT LLM to work efficiently inside this secure environment.
Read on developer.nvidia.comDetails
Nubank16 Sep 2026Guardrails
Designing Claude for ears, not eyes: An accessibility skill for blind and low-vision users
Problem Blind and low‑vision users struggled to use the chatbot because it was designed for visual interaction
Solution They added rules to the model’s instructions so it speaks clearly and follows screen‑reader conventions, making the assistant usable by audio‑only users
Read on building.nu.comDetails
Medium19 Aug 2026Guardrails
How we think about text classification in the LLM era
Problem The old NSFW classifier missed inappropriate content, showing unsuitable stories to users.
Solution Medium rebuilt the NSFW model using Snowflake's LLM inference, keeping data private and updating the pipeline without retraining on their own servers.
Read on medium.engineeringDetails
Replit17 Aug 2026Guardrails
Black-box pen tests on Replit
Problem Builders needed fast, affordable security checks for their apps before launch.
Solution They added black‑box penetration tests that run in a sandbox, simulating an attacker’s browser actions while also running a white‑box scan on the code.
Read on replit.comDetails
Hugging Face18 Jun 2026Guardrails
MosaicLeaks: Can your research agent keep a secret?
Problem Research agents unintentionally reveal private data through their web queries.
Solution They built a privacy‑aware RL training method that teaches agents to avoid leaking information while still answering multi‑hop questions.
Leakage reduced from 34% to 9.9%
Read on huggingface.coDetails
Shopify15 Jun 2026Guardrails
Teaching Sidekick to say no: automated data curation with LLM judge consensus
Problem The assistant never learned to refuse impossible merchant requests, leading to bad answers.
Solution Created an automated pipeline where several LLMs vote on whether a query should be refused, generated refusal examples, and fine‑tuned the model with this data.
Read on shopify.engineeringDetails
Dropbox12 Jun 2026Guardrails
How Dropbox uses MCP and Dash to close the design-to-code security gap
Problem Security requirements often get lost between design review and code review, making it hard to verify compliance.
Solution They built a system that uses a model‑context protocol, large language models, and Dropbox's Dash search to automatically pull relevant threat models into code reviews and check code against those requirements.
Read on dropbox.techDetails
Razorpay9 Jun 2026Guardrails
From 750 Hours to 2 Hours: AI-Powered Security Triage at Razorpay
Problem Developers were flooded with thousands of security alerts, most of which were false positives.
Solution They created an AI triage engine that reads full code context, decides if alerts are real, and auto‑remediates true issues, letting humans focus on real problems.
Triage time cut from 750 hours to 2 hours
Read on engineering.razorpay.comDetails
Replit22 Apr 2026Guardrails
Introducing Replit Auto-Protect
Problem Developers had to manually track security flaws and update libraries, which was slow and error‑prone.
Solution Replit Auto‑Protect scans projects for known CVEs, generates patches, and lets users apply them with two clicks through the Security Center.
Read on replit.comDetails
Replit21 Apr 2026Guardrails
Meet Replit Security Agent
Problem Pre‑launch security reviews took weeks, slowing releases.
Solution Created an AI Security Agent that automatically scans the whole codebase, builds a threat model, and reports vulnerabilities in under an hour using hybrid tools.
Full audit completed in under an hour
Read on replit.comDetails
Replit21 Apr 2026Guardrails
Defense in Depth: How Replit Secures Every Layer of the Vibe Coding Stack
Problem Developers worry AI‑generated code might be unsafe or break other users' environments.
Solution Replit layered isolation with hardened Linux containers, seccomp filters, continuous patching, and is moving to microVMs for stronger sandboxing, creating defense‑in‑depth.
Only one kernel exploit seen, no user impact
Read on replit.comDetails
Flipkart3 Mar 2026Guardrails
From Noise to Insight: How LLMs Are Transforming Threat Intelligence Analysis
Problem Analysts were overwhelmed by noisy, unstructured threat alerts and struggled to spot real risks.
Solution Created DeepInsight, an end‑to‑end system that gathers alerts, extracts key entities, feeds them to an LLM with a safety‑focused prompt, and stores concise, structured summaries for easy querying.
Read on blog.flipkart.techDetails
Flipkart15 Dec 2025Guardrails
Triksha - Securing AI with AI at Flipkart
Problem Prompt attacks let users trick AI support bots into unsafe or private answers.
Solution Flipkart created Triksha, a model that generates domain‑specific attack prompts to test and harden their guardrails, using pattern data, Markov chains and Gemini refinement.
14.6× higher bypass rate than existing guardrails
Read on blog.flipkart.techDetails