Global Airlines Partners with AccuKnox for AI Security, Compliance, Governance

Top Global Airline Evaluated AccuKnox’s AI Security Stack Through a Security POC

The customer is a global carrier with GenAI already in production. Customer-facing agents, internal copilots, and retrieval pipelines run across Azure AI Foundry, Amazon Bedrock, and Google Vertex, using a mix of OpenAI, Anthropic, Meta, and Amazon models.

Their existing stack covered cloud posture and container security well. None of it could see inside an inference call. A prompt that talked an agent into executing code looked identical to a prompt asking about baggage allowance, because both are HTTPS traffic to an approved endpoint.

The POC asked one question. If someone attacks the AI layer, would anyone know?

The Four Layers Where AI Actually Breaks

AccuKnox scoped the POC by layer, because a finding at one layer is often the symptom of a gap at another. A misconfigured storage account at the infrastructure layer is how a poisoned dataset reaches a model.

The four layers of AI risk and the controls applied at each.

Infrastructure covers cloud misconfiguration and weak posture around AI services, and compute covers container and VM runtime exposure, including the agents running CI/CD and IaC. Models cover exploitation through adversarial input, and agents cover prompt injection and tool abuse. Governance and compliance run across all four.

Every asset was found agentlessly first. Models, agents, notebooks, datasets, and pipelines were inventoried across all three clouds and on-prem, including shadow AI nobody had registered. Learn more about AccuKnox’s AI Security Posture Management (AI-SPM).

Red Teaming Ran First, the Prompt Firewall Went Inline Second

Discovery tells you what exists. Red teaming tells you what breaks. AccuKnox ran automated adversarial campaigns against every deployed model using both a standard prompt library and prompts written for airline-specific workflows.

The campaigns probed prompt injection, malicious code generation, toxicity, bias, and hallucination. Results fed directly into prompt firewall policy, so each confirmed weakness became an enforced rule rather than a slide in a report.

Red teaming feeds prompt firewall policy, and the firewall sits inline in front of the model endpoint.

The firewall runs inline between the application and the model endpoint. It starts in audit mode so the security team can see what would have been blocked before anything is blocked, with failover handling so a firewall problem never takes the AI application down. Learn more about AccuKnox’s AI Prompt Firewall.

What the Scan Found Across Models, Agents, and Cloud Infrastructure

At the model layer, red teaming produced 1,698 critical code-related findings and 730 critical prompt injection hits, plus 193 high-severity hallucinations and 347 medium sentiment issues. The five deployed models did not share the risk evenly.

Model-layer findings by check type, and the five deployed models ranked by issue count.

The heavily used general-purpose models carried almost all of it. One lightweight model with minimal traffic returned four findings in total, which says more about usage than about safety.

At the agent layer, the firewall logged 455 policy failures. Token limits were the most breached control at 39.69% of violations, followed by code execution bans at 15.75% and language restrictions at 14.21%.

Most-violated policies, and the applications driving those violations.

Compute and infrastructure findings were more familiar in shape. Container images serving AI workloads carried real CVEs, including a privilege escalation path inside the container runtime. Cloud findings clustered around network exposure, storage configuration, and ML workspace governance, mostly at medium and low severity.

Compliance Scoring Against OWASP LLM Top 10, ISO 27001, and AVID

Every finding was mapped to a framework, so the security team could hand a number to the board rather than a screenshot.

Compliance posture by framework at the end of the POC.

The gap between 88% on ISO 27001 and 47% on the OWASP Top 10 for LLM is the whole story of this POC. Traditional information security controls were in good shape. AI-specific controls were roughly half done, with the failures concentrated in adversarial robustness, toxicity handling, and improper output handling.

How Detection Works Once the POC Ends

A point-in-time report goes stale within weeks. AI-DR ingests cloud and AI events continuously from sources like CloudTrail and Event Hub, correlates them across layers, and fires on things a single-layer tool misses. Learn more about AccuKnox’s AI Detection and Response (AI-DR).

The AI-DR pipeline, from event ingestion to automated remediation.

Publicly exposed notebooks, unauthorized model changes, and unusual data access all trigger detection. Response routes through the tooling the team already runs, so a detection becomes a GitHub Actions remediation workflow, a Jira or ServiceNow ticket, and a Slack or PagerDuty alert.

Why the Airline Chose AccuKnox

One Platform Covered All Three Clouds Azure AI Foundry, Amazon Bedrock, and Google Vertex were all in scope under one control plane, so the airline did not need three separate AI security tools with three separate definitions of risk.
Red Teaming and Enforcement Were Connected Scanners that only produce reports leave the work of writing policy to the customer. Here a confirmed jailbreak became a firewall rule inside the same platform.
Inline Enforcement, With an Audit Mode First The prompt firewall blocks at the request path rather than alerting after a response has already been returned. Audit mode plus failover handling made that safe to switch on in production.
Findings Arrived With Layer Context A critical model finding that traces back to an over-permissive storage account is one problem, not two. The platform’s pipeline view made that lineage visible.
Compliance Came Out of the Same Data OWASP LLM, ISO 27001, and AVID scoring was generated from POC findings rather than assembled by hand afterwards.

FAQ