Anthropic Mythos lowers discovery costs, not the vulnerability burden

At a glance

What this is: This is an analysis of Anthropic Mythos and the finding that AI is accelerating vulnerability discovery more than it is changing attack mechanics.

Why it matters: It matters because security and identity teams still have to contain the blast radius after discovery, which keeps least privilege, segmentation, and access controls central to both human and non-human identity governance.

By the numbers:

• Only 44% of organisations have implemented any policies to manage their AI agents, despite 92% agreeing that governing AI agents is critical to enterprise security.
• Systems with least-privileged AI access had a 17% incident rate versus 76% for over-privileged systems, making poorly scoped AI access 4.5x more likely to lead to a security incident.

👉 Read Expel's analysis of Anthropic Mythos and AI-accelerated vulnerability discovery

Context

Anthropic Mythos is best understood as an acceleration layer for vulnerability discovery, not a new class of vulnerability. The primary security problem is that software weakness that was previously expensive to uncover can now be surfaced quickly at scale, which raises pressure on code review, dependency mapping, and containment controls across AI and software estates.

For IAM and NHI teams, the real question is not whether AI can find more flaws, but what happens after it does. If AI-assisted discovery is faster, then access boundaries, segmentation, and least privilege become the controls that determine whether a weakness becomes an incident, especially where service accounts, tokens, or AI systems hold broad access.

That makes this topic relevant beyond application security. It touches the governance of human access, machine identity, and AI-assisted development workflows, which is why a strategy conversation at executive level is more useful than waiting for a single technical fix.

Key questions

Q: How should security teams respond when AI discovers vulnerabilities faster than humans can patch them?

A: They should shift from point-in-time vulnerability handling to continuous exposure reduction. That means prioritizing the exploitable paths an attacker can chain now, not only the highest-severity findings, and tying remediation to identity controls, segmentation, and blast-radius reduction. If an AI attacker can move faster than the patch cycle, containment becomes the primary control objective.

Q: Why do segmentation and least privilege still matter in an AI-driven threat landscape?

A: They still matter because AI speeds up discovery, not the physics of compromise. Once an attacker has a foothold, segmentation limits where they can move and least privilege limits what they can do. Those controls turn a fast-discovered flaw into a contained event rather than a broad environment-wide incident.

Q: What do organisations get wrong about vulnerability discovery?

A: They often treat discovery as proof of risk. Discovery only says something exists, not that it can be exploited or chained into impact. Security teams need validation that tests reachability, privilege paths, and business consequence, otherwise remediation time is wasted on theoretical issues.

Q: Who should own strategy when AI changes attacker speed?

A: Ownership should sit with security leadership, engineering leadership, and risk leaders together, because AI-assisted discovery affects budgets, tooling, and recovery assumptions at the same time. If one team treats it as just an operations issue, the organisation will underinvest in the controls that actually reduce impact.

Technical breakdown

Why AI-assisted vulnerability discovery changes the attacker economics

Mythos matters because it shifts the economics of discovery. Many software weaknesses have always existed, but they remained low priority because finding and chaining them required specialist time, tooling, and patience. AI compresses that work into a much shorter cycle by helping researchers enumerate likely flaws, test combinations, and prioritise promising paths. The security issue is not magic exploitation. It is a lower cost to reach the same result, which increases the volume of issues attackers can realistically pursue.

Practical implication: security teams should assume more findings will be surfaced faster and shift remediation triage toward the most exposed paths first.

Why source code review becomes a defender advantage

Defenders often have an information advantage that attackers do not: direct access to source code, dependency manifests, and internal build logic. That makes AI-assisted review more effective on the defensive side because models can inspect the actual implementation instead of reverse engineering binaries. This does not replace human review, but it can improve the speed of identifying patterns such as insecure input handling, secret exposure, or unsafe trust assumptions in code and pipelines. The gap is less about capability than adoption and workflow integration.

Practical implication: point LLM-assisted analysis at owned source code, SBOMs, and build artefacts before adversaries do.

Why segmentation and least privilege still govern impact

AI may change discovery speed, but it does not change post-exploitation physics. Once an attacker gets in, they still need to move across systems, reach data, and escalate to meaningful access. Segmentation reduces reach, least privilege limits what each account or service can do, and phishing-resistant MFA makes privilege pivoting harder. That is why traditional controls remain relevant in an AI-accelerated threat landscape. The architecture still determines whether a discovered flaw becomes a contained issue or a broad compromise.

Practical implication: measure blast-radius controls, not just vulnerability counts, because containment is now the decisive layer.

NHI Mgmt Group analysis

AI-assisted discovery exposes a security debt that many organisations already carry. Mythos does not introduce novel weakness so much as reveal how much unresolved weakness already exists in legacy software, exposed dependencies, and weak review processes. The important governance issue is that discovery has outpaced remediation planning. Practitioners should treat this as an acceleration of existing risk, not a new threat category.

Defenders have a real structural advantage when they can analyse their own source code directly. Attackers working from binaries, public artefacts, or externally visible behaviour face a harder problem than internal teams reviewing source, dependency graphs, and CI artefacts. That advantage only matters if organisations operationalise it. In practice, AI review should be built into secure development workflows rather than treated as an occasional experiment.

Blast-radius control: the real contest is no longer vulnerability discovery, but how much damage a discovered flaw can do. The article reinforces that segmentation, MFA, and least privilege still decide whether an exploit becomes a minor exposure or a major incident. This aligns with NIST-CSF, NIST-800-53, and MITRE-ATT&CK thinking on containment and access control. Practitioners should measure exposure by reachable privilege and segmentation depth, not just by scanner output.

Executive alignment is now a security control, not a soft governance exercise. The article’s strongest operational point is that defenders need shared strategy before the next acceleration wave arrives. That means budget, tooling, and ownership for AI-assisted security work cannot sit in separate silos. For identity programmes, this is especially important where human access, service accounts, and AI systems intersect in the same workflow.

AI changes the tempo of discovery, but not the rules of access. The fact that segmentation and least privilege remain effective is itself the signal: identity governance still determines whether faster discovery leads to broader compromise. That makes this a governance problem as much as a technical one, and practitioners should fold AI into existing identity and access control models rather than building parallel exceptions.

What this signals

Blast-radius control: AI-assisted discovery is likely to make vulnerability management feel noisier, but the programme outcome that matters is still exposure reduction. Teams that can tie scanning, source review, and identity controls together will handle faster discovery without expanding operational risk.

The identity signal is clear. As AI systems gain broader access in production workflows, the same access review logic used for human accounts becomes insufficient unless it is extended to service accounts, tokens, and AI tooling. That is why practitioners should treat machine privilege as part of the same governance perimeter as human access.

Security leaders should also expect the conversation to move from tool capability to operating model. The practical question is no longer whether AI can find flaws, but whether the organisation can fund and govern the remediation path quickly enough to matter.

For practitioners

• Build AI-assisted source review into secure development Use LLM-assisted analysis on owned source code, dependency manifests, and CI artefacts to identify exposed secrets, unsafe trust boundaries, and weak input handling before attackers do.
• Prioritise containment over raw defect counts Rank remediation by reachable privilege, segmentation gaps, and externally exposed paths so the first fixes reduce blast radius rather than simply lowering scanner volume.
• Reassess identity controls for AI-enabled workflows Check whether service accounts, API keys, and AI tooling have broader access than the human roles they support, then tighten least privilege where machine and human access overlap.
• Align executive strategy around AI attack tempo Set shared ownership for AI-assisted security work across security, engineering, and risk so budget, tooling, and remediation priorities move at the same pace as attacker capability.

Key takeaways

• Mythos accelerates the discovery of old weaknesses, which means exposure management now matters more than assuming obscurity will protect bad code.
• The most useful defender advantage is direct access to source, dependency data, and build artefacts, because those inputs let AI work on what attackers cannot easily see.
• Segmentation, MFA, and least privilege still determine whether an AI-discovered flaw becomes a limited event or a major compromise.

Key terms

• Security by obscurity: Security by obscurity is the informal reliance on information being hard to find rather than properly protected. It can mask poor governance for a while, but it fails quickly once search, automation, or AI makes hidden content easy to discover and retrieve.
• Blast Radius: The potential scope of damage if a specific credential or identity is compromised. Identities with broad permissions have a larger blast radius and represent a higher priority for least-privilege enforcement and security controls.
• Least Privilege: A security principle requiring that every identity — human or non-human — is granted only the minimum permissions necessary to perform its function. Least privilege is the single most effective control for reducing NHI blast radius.
• AI-scale vulnerability discovery: The use of AI to identify weaknesses across applications, identities, integrations, and workflows at a speed that can exceed manual review. The security challenge is not discovery itself, but whether the organisation can close the identity paths it exposes.

What's in the full article

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

• James Shank's full commentary on why Mythos changes attacker tempo more than attacker capability
• The CSA paper context that Expel discusses, including the idea of a Mythos-ready security program
• The article's explanation of where defenders can use source-code access to offset AI-assisted discovery
• The leadership argument for aligning executive strategy before AI-driven attack speed increases further

👉 Expel's full post expands on the defender advantage, the strategy discussion, and the security fundamentals that still matter.

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 identity and security practitioners connect access control to real-world operational risk.