Topic starter
02/08/2026 11:48 am
TL;DR: AI models will speed up attack workflows, but their effectiveness depends on deep system context, according to Aikido’s analysis of Mythos and 1,000 real-world AI penetration tests. The defender advantage still comes from source code, runtime visibility, and internal logic, not raw model capability.
Editorial analysis by NHI Mgmt Group, based on content published by Aikido: “The cybersecurity doomerism around Mythos doesn't match what we see on the ground”.
Key questions
A: Use AI models as high-signal assistants, not as complete scanners.
Q: Why do whitebox tests produce better results than greybox tests for AI-assisted appsec?
A: Whitebox tests expose the implementation details that let an AI reason about hidden branches, vulnerable flows, and unsafe assumptions.
Q: What signs show that an AI security workflow is missing enough context?
A: The main signs are shallow findings, inconsistent severity ratings, and weak explanations for why a weakness matters.
Practitioner guidance
- Combine source and runtime visibility Correlate source code review, runtime telemetry, and dependency metadata before trusting AI-assisted findings on production apps.
- Prioritise whitebox testing for high-value applications Use fuller access for applications where business logic, internal APIs, and credential flows materially affect risk.
- Govern dependency changes as security events Treat unexpected package updates, maintainer-account compromise, and transitive dependency changes as review triggers, not routine hygiene.
Bottom line: AI model capability does not automatically create attacker advantage in application security because context remains the limiting factor.
Explore further
View Full Forum → | NHI Foundation Course → | Our Services → | Read the full analysis →
02/08/2026 12:58 pm
Context is now the controlling variable in AI-assisted application security: model capability does not produce reliable security outcomes unless the testing system can see code, runtime behaviour, and internal logic together. That is why the same model can look impressive in lab conditions and underperform in constrained environments. For security teams, the real question is not how strong the model sounds, but how much of the system it can actually reason over.
A question worth separating out:
Q: How do NHI credentials change the way AI should be used in appsec?
A: NHI credentials add a trust layer that AI cannot infer reliably from static code alone. Service accounts, tokens, and API keys shape which actions are actually possible, so the analysis has to include issuance, scope, and runtime use. If those identities are invisible, AI testing will miss the real governance boundary.
👉 Read our full editorial: AI model context matters more than attacker hype in appsec