Mythos and Glasswing show why AI-powered pentesting is shifting

At a glance

What this is: This analysis argues that Mythos and Project Glasswing changed urgency and expectations around AI-assisted security testing, not the underlying reality that exposed services, legacy integrations, and validation controls remain the real weakness.

Why it matters: It matters because IAM, PAM, and security teams still need governed access, trusted execution, and accountability around offensive automation, especially when AI tools can probe live environments at scale.

By the numbers:

• Only 5.7% of organisations have full visibility into their service accounts.
• 96% of organisations store secrets outside of secrets managers in vulnerable locations including code, config files, and CI/CD tools.
• 79% of organisations have experienced secrets leaks, with 77% of these incidents resulting in tangible damage.

👉 Read Ethiack's analysis of Mythos, Project Glasswing, and AI-powered pentesting

Context

AI-assisted pentesting changes the speed and breadth of attack simulation, but it does not remove the underlying governance problem: organisations still need trustworthy boundaries around what is tested, how findings are validated, and who is accountable for the results. In practice, the security gap is less about whether a model can find issues and more about whether the enterprise can operationalise those findings safely. For identity teams, that means the control plane around credentials, service accounts, and privileged workflows remains central.

The article frames Mythos and Project Glasswing as a market signal that security testing is moving toward higher automation and broader coverage. That is plausible, but the more durable issue is the attack surface itself: exposed services, forgotten integrations, and weak credential hygiene still dominate real compromise paths. The starting position described here is common in modern enterprises, not unusual.

Key questions

Q: How should security teams govern AI-assisted web testing tools?

A: Treat AI-assisted testing as a governed workflow, not a convenience feature. Define which targets, data, and actions the tool may touch, assign separate credentials and logs, and require human approval for anything that could affect production systems. The goal is to keep the agent’s scope narrow enough that its actions remain attributable, reviewable, and reversible.

Q: Why do exposed services and leaked credentials still matter when AI can find vulnerabilities faster?

A: Because faster discovery does not remove weak access paths. Exposed services, stale secrets, and over-privileged accounts remain the shortest route from reconnaissance to compromise. AI can increase the speed of finding those weaknesses, but it does not replace credential hygiene, access control, or monitoring. The attack surface still decides the outcome.

Q: What do researchers get wrong about using AI in offensive security?

A: The common mistake is treating AI as a substitute for verification. AI can accelerate recon, pattern matching, and report writing, but it cannot confirm that an exploit works or that a claimed vulnerability exists. If researchers do not independently validate their findings, they are producing speculation, not professional disclosure.

Q: Who is accountable when AI pentesting is run outside approved scope?

A: Accountability should be defined before the pilot starts. Security owns authorisation and controls, while procurement, privacy, and legal must sign off on data handling, retention, and liability boundaries. If the test crosses scope, the absence is usually governance, not just tooling.

Technical breakdown

How AI-powered pentesting works against exposed attack surfaces

AI-assisted pentesting combines code analysis, prompt-driven reasoning, and tool use to search for vulnerabilities faster than a manual workflow. In practice, the model may infer likely weak points, chain issues across components, and generate exploitation hypotheses even when source code is limited. That changes throughput, not the laws of exposure. A model can only amplify what exists in the environment, including weak secrets handling, over-permissive access, and neglected third-party integrations. The operational challenge is separating promising findings from noise and ensuring offensive actions stay bounded within approved scope.

Practical implication: treat AI-driven testing as an execution layer that still needs strict scope controls, audit logging, and human validation.

Why validation and accountability matter more when findings arrive faster

The article points to a growing gap between model output and enterprise security operations. That gap exists because vulnerability discovery is only useful when findings are trusted, prioritised, and tied to remediation ownership. AI can accelerate discovery, but it cannot by itself establish proof, severity, or business impact. This is where governance matters: validated evidence, repeatable triage, and clear handoff into remediation workflows determine whether automation reduces risk or simply increases alert volume. For identity-linked attack paths, weak credential governance can turn a fast discovery into a fast compromise.

Practical implication: build a verification and ownership workflow before scaling AI-generated findings into production response.

MCP, terminal access, and the expansion of agentic attack tooling

When AI systems can connect to MCP tools, the terminal, and scripted skills, they move from passive analysis to active task execution. That makes the surrounding identity model more important, not less. Each tool connection expands the set of credentials, permissions, and approvals that govern the system’s actions. Without clear segmentation, an AI-driven workflow can inherit excessive privilege or operate outside intended guardrails. This is especially relevant where offensive tooling touches cloud accounts, secrets stores, or privileged automation paths that were never designed for autonomous decision-making.

Practical implication: govern AI tool access with least privilege, short-lived credentials, and explicit approval boundaries.

NHI Mgmt Group analysis

AI-powered pentesting exposes a governance gap, not a security reset. The article is right that model capability is improving, but the underlying enterprise problem is still validation, scope, and accountability. Offensive automation only becomes operationally useful when organisations can trust the output and assign remediation ownership. The practical conclusion is that security teams need a control plane for AI-assisted testing, not just better models.

Identity and secrets hygiene remain the deciding variables in AI-enabled attack paths. The most realistic failure mode is not a model magically breaking everything, but a fast path through exposed services, leaked secrets, and over-privileged access. That intersection is where NHIs matter most, because service accounts, API keys, and tokens are the bridge between discovery and exploitation. In NIST-CSF terms, access control and continuous monitoring still determine whether the attack stops at discovery or reaches impact.

Project Glasswing signals a market shift toward broader offensive coverage, but breadth alone is not governance. Automated testing that focuses on audited software can still miss the real risk sitting in obscure integrations and production-only dependencies. That is a useful reminder that security programmes fail when they overfocus on visible assets and under-govern the hidden trust relationships that tie environments together. The practitioner takeaway is to align testing with exposure, not just with code volume.

Named concept: validation latency. This article highlights the delay between machine-generated findings and enterprise-grade confidence in those findings. The longer that gap, the more likely organisations are to waste cycles on noisy output or miss the window for containment. For practitioners, shortening validation latency becomes a measurable security objective, especially in AI-assisted offensive workflows.

Agentic offensive tooling expands the attack surface of the testing process itself. Once AI systems can chain tools and act across environments, the security of the testing workflow becomes part of the control objective. That creates a new governance requirement around delegated authority, auditability, and scope enforcement. The conclusion for practitioners is simple: secure the tester as carefully as the target.

What this signals

AI-assisted pentesting will increase the number of findings organisations can generate, but it will also widen the gap between discovery and accountable remediation unless validation workflows mature. The practical signal for security programmes is to invest in evidence handling, ownership mapping, and control validation before scaling autonomous or semi-autonomous offensive tooling.

Validation latency: enterprises should measure how quickly model-generated findings become trusted decisions, because slow validation turns automation into backlog. That latency is now a programme risk, not just an operational inconvenience.

For practitioners

• Define scope boundaries for AI-assisted testing Require explicit target lists, forbidden actions, and escalation rules before any model is allowed to probe production-connected systems. Tie those rules to logging and review so offensive actions remain attributable.
• Validate AI-generated findings before triage Create a verification step that separates likely issues from exploitable evidence, then route confirmed findings into existing remediation queues with named owners and due dates.
• Harden service account and token hygiene Prioritise the credentials that connect testing tools to cloud consoles, terminals, and CI/CD systems. Rotate exposed secrets, remove standing privileges, and track every non-human identity that can reach production.
• Review tool access for agentic workflows Map every MCP connection, CLI integration, and automation hook used by AI testing systems, then apply least privilege and short-lived credentials to each path.
• Measure validation latency Track the time between a model-generated finding and a confirmed, actionable security decision. Long validation delays usually indicate weak ownership, noisy inputs, or a broken triage workflow.

Key takeaways

• AI-powered pentesting increases speed and coverage, but it does not remove the need for scope, validation, and accountability.
• Service accounts, secrets, and standing access remain the real bridge between machine-generated findings and real compromise.
• Security teams should govern the testing workflow as carefully as the target environment, especially when tools can act across MCP, terminals, and automation hooks.

Key terms

• AI-Driven Pentesting: AI-driven pentesting uses reasoning systems to plan and execute multi-step attack simulations against applications or infrastructure. It differs from rule-based scanning because it can follow workflows, track state, and evaluate whether multiple weaknesses combine into a viable compromise path.
• Validation Latency: The time between a weakness being introduced, proven exploitable, remediated, and confirmed closed. Shorter validation latency means defenders can trust that their controls are keeping pace with release and attack speed.
• Agentic Tooling: Software that can move beyond answering prompts and take actions through connected systems. In identity terms, it behaves like a delegated non-human actor because its risk depends on the permissions, integrations, and execution paths it can use at runtime.
• Model Context Protocol: Model Context Protocol is an open protocol that lets AI agents connect to tools and data sources. It expands what an agent can reach, so governance has to cover not only the model and its prompts, but also every system that can receive or return agent-driven data.

What's in the full article

Ethiack's full blog covers the operational detail this post intentionally leaves for the source:

• Practical examples of how AI-assisted offensive testing is structured across real environments and workflows
• The vendor's view of how model capability changes attack simulation depth, coverage, and response planning
• Operational context on where AI findings still need human judgement, prioritisation, and validation
• Discussion of how teams can balance offensive automation with scope control and accountability

👉 The full Ethiack post expands on the market reaction, attack-surface discussion, and security industry implications.

Deepen your knowledge

The NHI Foundation Level course, the industry's only accredited NHI security programme, covers NHI governance, machine identity security, and secrets management. It helps security practitioners connect access control, lifecycle management, and operational accountability across identity programmes.