What are the opening hours?

Monday 9am - 5pm Tuesday 9am - 5pm Wednesday 9am - 5pm Thursday 9am - 5pm Friday 9am - 5pm

Prediction Guard

Name: Prediction Guard
Address: 101 N 4RTH Street, Lafayette, IN, 47901, US
Telephone: +17654095758

Home
United States
Lafayette, IN
Prediction Guard

Scale autonomous AI agents with real control! 🚀 Self-hosted runtime controls that keep your agent fleet in-boundary and accountable.

10/09/2026

It's 2:14 AM. A payments reconciliation agent starts pulling customer account records it has never touched before. Within seconds, a runtime control flags it, the agent is quarantined, and a kill switch cuts its access.

The control worked. Now comes the hard part.

By 8:00 AM, the CRO, the CISO, internal audit, and eventually an examiner will all ask the same question: why?

For AI leaders in financial services, stopping a bad agent is table stakes. Explaining the stop, with evidence that holds up to a regulator, is the real test.

In her latest post, Katie Bowen breaks down why agent explainability is now a board level issue, the three questions every investigation has to answer, and how agent identity plus tracing turns a 2:14 AM quarantine into a clean answer by morning.

Read the full post in the comments.

10/08/2026

Prediction Guard is headed to Midwest AI Summit on October 15. Find us in the Engineering Lounge and the expo area, we'll have team members there that you can sit down with one-on-one to work through your own agent use cases, architecture questions, or whatever you're currently stuck on.

If you're building or securing AI agents and want real time with people doing this work daily, this one's worth carving out a day for.

Oct 15 | Indianapolis
Register in the comments below

10/06/2026

Nvidia just launched something called the Open Agent Safety Platform, built with over 100 industry partners, designed specifically to rapidly quarantine rogue AI agents.

When a chip maker is building hardware-and-software security specifically to isolate misbehaving agents, that's worth sitting with. The question of how you actually stop an agent isn't theoretical enough anymore to leave to individual application teams. It's becoming infrastructure.

Quarantine is really another word for something we've been saying for a while: a kill switch has to be real, tested, and ready, not a line in an incident response doc nobody's rehearsed. The specific mechanism matters less than the convergence itself. Enterprises, labs, and now chip makers are all landing on the same conclusion. Agents need a way to be stopped that doesn't depend on their own willingness to stop.

That's the actual shape of operational control of agents: scoped access, a defined blast radius, and a kill switch that works the moment it's needed, whether it lives in your own control plane or in the hardware stack underneath it.

10/01/2026

Aishwarya and Katie are heading to InfoSec World in Orlando this month.

Catch Aishwarya's talk: "Can Leaderboards Predict Production Safety?"
Tuesday, Oct 13, 11:50 AM–12:30 PM

Her research (BenchRisk, NeurIPS 2025) applied NIST's risk management framework to 26 widely used LLM benchmarks and found 57 failure modes that can lead teams to the wrong conclusions about model safety. A benchmark score that says "safe" can still hide gaps in coverage, reproducibility, or data contamination, gaps that turn into real vulnerabilities once a model hits production.

Having evaluated 30+ open-weight models for regulated deployments, her core finding: bigger isn't always better. A leaderboard leader can lose to a smaller model on your specific task, your hardware, or a risk no benchmark was built to catch.

If you're there, let's connect! Book a meeting in the comments below.

09/28/2026

Prediction Guard is now part of the Docker Verified Publisher agentic program, and our kit, predictionguard/sbx-predictionguard, is live on Docker Hub.

Here's why that matters. Agent frameworks let you stand up an agent in an afternoon. The hard part is everything after that: proving which identity an agent acted under, giving it least agency, keeping model traffic and sensitive data inside your boundary, and producing an audit trail your security team will actually accept. That's where most prototypes stall.

The kit handles that with two gates of defense.
Gate one is the sandbox. Network egress is restricted to a single endpoint, and credentials never enter the VM. If an agent reaches for something it shouldn't, the blast radius stays contained.

Gate two is our self-hosted control plane. Every model call and MCP tool call passes through it, with auditable, deterministic controls applied to live agent behavior rather than after the fact.

What you get:
- MCP tool scoping per agent identity, a built-in MCP proxy, logging of every tool call, and configurable human-in-the-loop approvals
- Runtime controls for PII redaction, credential scrubbing, and injection detection, plus behavior controls for memory poisoning, runaway token use, and goal drift
- OpenAI, Anthropic, MCP, and agent SDK compatible endpoints, so LangGraph, Pydantic AI, CrewAI, OpenCode, and Claude Code work with zero refactoring
- Signed images, Docker Scout scanning, and the Docker Verified Publisher badge

It runs wherever you already run workloads: on premises, air-gapped, or in a private VPC. The controls map to NIST, AIUC-1, OWASP, ISO, and EU AI Act guidance.
One command to launch a governed agent sandbox. Links to the kit and a guided evaluation are in the comments.

Thanks to Docker, Inc. and Eli Aleyner for the partnership, and to Daniel Whitenack for leading the way.

09/25/2026

On September 3rd, two members of Congress introduced legislation citing direct quotes from AI agent messages as evidence for why it's needed. Not a hypothetical, not a whitepaper example, actual agent-to-agent chatter from a real incident, read into the legislative record.

That's a preview of where this goes next. Agent behavior is becoming evidence. Not evidence you curate for a press release, evidence a regulator, a plaintiff's attorney, or a senator's staffer pulls directly from whatever logs happen to exist.

Most organizations running agents today couldn't produce that record if asked. Not because they're hiding something, but because the logging that exists was built for debugging, not for standing up under outside scrutiny.

The organizations that will handle this moment calmly are the ones who already treat every agent action as something that needs to be provable after the fact, not just observable while it's happening. That's not a legal team's problem to solve after an incident. It's an architecture decision made before one.

09/23/2026

Prediction Guard is on the road this week! Sharan is speaking at We Are Developers in San Jose, with Steve also there.

Catch his talk: "Two Gates of Defense: Running AI Agents Safely with Docker Sandbox and Prediction Guard"
Thursday, Sep 24, 3:15-3:30 PM PDT | Docker Pavilion, San Jose Convention Center

Two layers of defense, one live agent, and a poisoned tool output on purpose. Come see what breaks, and what holds.

Want to grab time with Steve while he's there? Link to book in the comments 👇

09/21/2026

We had such a good time at Midwest AI Summit last year, and we're genuinely excited to see what everyone's been building in the AI space since then.

This year's event is October 15 in Indianapolis, and if you haven't grabbed your seat yet, use code PredictionGuard20 for 20% off your ticket.

Come say hi, swap notes on what's actually working in production, and let's see what this community has been up to.

Oct 15 | Indianapolis
Register in the comments below

09/18/2026

"We have AI governance" is quietly becoming the "we take security seriously" of 2026. Everyone says it. Almost nobody can show you what it means.

Ask around your own org. What does "we have governance in place" actually point to? You'll get a steering committee. A policy document. A model inventory. A quarterly risk review. Someone in legal who signs off.

Every one of those is real. And not one of them does anything the moment an agent reaches for a tool it shouldn't touch.

Gartner's own hype cycle makes the point without meaning to. Trace the curve to its end, and the only thing sitting alone on the Plateau of Productivity is Privacy Management Tools. That's not a scoring error. That's what a governance chart matures into when the word was built for something that holds still.

Governance, as most companies practice it, describes intent. It tells an auditor what you meant to do. It doesn't sit in the path of a live agent and stop it. And an agent in production doesn't run on intent, it runs on whatever it's actually allowed to do, right now, at the moment it acts.

Catherine Bowen's piece traces why this keeps happening and argues the vocabulary of 'AI governance' hasn't caught up to the agent era. Worth reading if you're running agents anywhere near production.

Link in the comments. And a fair gut-check for your own stack: if something went wrong with one of your agents yesterday, could you point to the moment it was stopped, or only to the policy that said it shouldn't have happened?

09/14/2026

Missed our webinar with Carahsoft on Zero Trust for AI Agents? We covered what federal agencies and contractors need to do now to keep agentic AI systems auditable and standards-aligned. The full replay is up, link in the comments 👇

Address

101 N 4RTH Street
Lafayette, IN
47901

Opening Hours

Monday	9am - 5pm
Tuesday	9am - 5pm
Wednesday	9am - 5pm
Thursday	9am - 5pm
Friday	9am - 5pm

Telephone

+17654095758

Website

http://www.predictionguard.com/

Alerts

Be the first to know and let us send you an email when Prediction Guard posts news and promotions. Your email address will not be used for any other purpose, and you can unsubscribe at any time.

Contact The Business

Send a message to Prediction Guard: