Deflected Platform

AI Incident Response

When your AI is the breach.

On-call expert response for when an AI system is breached, manipulated, or leaking. Deflected provides containment, forensics, and recovery on a standing retainer.

The risk this closes

When a model is jailbroken, poisoned, or quietly leaking, your standard incident playbooks don't apply — the evidence lives in prompts, embeddings, and model behavior, not just logs. Every hour of dwell time widens the blast radius while your team improvises against an unfamiliar failure mode.

AI incidents don't look like the breaches your team trained for. A prompt-injection attack turns trusted input into an instruction the model obeys. A manipulated or poisoned model behaves normally until the exact moment an attacker needs it not to. And sensitive data can walk out through the model's own output — no exfiltration path, no malware, just a well-crafted request. There is no packet capture for a jailbreak and no antivirus for a poisoned weight.

A generic incident-response team can secure the infrastructure around a model but rarely knows how to interrogate the model itself — how to reconstruct an attack from a prompt trail, distinguish manipulation from expected behavior, or tell whether a system is safe to bring back online. Meanwhile the clock runs: an AI system left compromised keeps making decisions, answering users, and touching data with every request. The faster containment, forensics, and recovery begin, the less an incident costs you.

Built to run in production

1

Contain

Isolates the affected model, pipeline, or agent to stop the bleeding without halting the business.

2

Investigate

Traces root cause through prompts, data, and model behavior to establish what actually happened.

3

Recover

Restores trusted state, closes the entry path, and hardens against a repeat of the same attack.

4

Retain

Keeps our responders on standby under a defined SLA so help is minutes away, not days.

What you get

Rapid containment

Expert responders isolate the compromise fast to limit exposure and downtime.

Forensic root cause

A defensible account of how the incident happened and what data was affected.

Retained response SLA

A standing retainer with guaranteed response times so you're never facing it alone.

Who we stand behind

AI Incident Response is built for the teams who own an AI system when something goes wrong — and need expertise they don't have to build in-house before the alarm sounds.

Security and IR teams

Capable responders who own the perimeter but lack deep AI expertise. We become the specialists on call when the incident is the model itself.

Companies running AI in production

Any organization with models, agents, or LLM features touching real users and real data — where a compromise has immediate operational and legal weight.

Executives who need a safety net

Leaders accountable for AI risk who want a retained response capability in place — so the answer to "who handles this?" is settled before it's ever asked.

In the real world

Three ways an AI system fails — and how a retained response changes the outcome.

Prompt injection in production

A live assistant starts ignoring its guardrails and following attacker-planted instructions. We isolate the affected path, trace the injection to its source, and close it without taking the product offline.

Data exfiltration via output

You suspect the model is surfacing records, secrets, or PII it should never expose. We reconstruct what was asked and returned, scope the exposure, and give you a defensible account of what left.

A manipulated or poisoned model

A model in use is behaving wrong in ways that point to tampered training data or weights. We confirm whether it was manipulated, quantify the impact, and restore a trusted version.

Why teams retain us

Incident response for AI is its own discipline. This is what sets our engagement apart.

AI-native forensics

We investigate where AI incidents actually live — prompts, embeddings, training data, and model behavior — not just the logs and endpoints around them.

Retained, pre-arranged SLA

A standing retainer means the relationship, access, and priority response are set up in advance. When it hits, we're already your team — no procurement, no cold start.

End to end, not just triage

Containment, root-cause forensics, and recovery in one engagement — we stop the incident, explain it, and get you back to a trusted state.

Quantum-secured by default

Every byte handled by AI Incident Response is protected with NIST-standardized post-quantum cryptography — ML-KEM-1024 key encapsulation (FIPS 203) in a hybrid scheme. Your data stays sealed against harvest-now, decrypt-later attacks, today and after quantum computers arrive.

Questions, answered

How does the retainer work?
You engage us on a standing retainer, so the relationship is in place before an incident happens. Up front we map your AI systems, agree on how you reach us, and pre-arrange the access and authority our responders need. When something goes wrong, you invoke the retainer and we begin immediately — no procurement cycle, no onboarding from scratch, no cold start while the clock runs.
What is the response time?
Retainer clients get priority response under a defined SLA agreed at the start of the engagement. Because the relationship, context, and access are established in advance, we can move the moment you invoke it rather than spending the first hours getting oriented. We set specific response targets with you based on your risk profile and the criticality of the AI systems in scope.
What does an engagement cover?
A full response cycle: containment, forensics, and recovery. We isolate the affected model, pipeline, or agent to stop the damage without halting your business; investigate root cause through prompts, data, and model behavior to establish exactly what happened and what was exposed; and restore a trusted state, close the entry path, and harden against a repeat. You come away with a defensible account of the incident and a system safe to run again.
Do we need this if we already have an IR team?
Yes, in most cases. A strong IR team secures the infrastructure around a model, but AI incidents live inside the model — in prompts, embeddings, training data, and behavior that conventional tooling doesn't capture. We work alongside your team as the AI specialists, handling the parts that require deep model expertise while they run the broader response. You keep ownership; we bring the capability that's hard to staff for full-time.

Put a response plan in place before you need it

The worst time to find an AI incident responder is during the incident. Book a working session — we'll map your AI systems, define your retainer, and make sure help is pre-arranged the day something goes wrong.