Public Benchmarks
These are the results of measuring how IAM Privilege Escalation Detection finds and judges paths, using public material that anyone can rerun the same way. The tables on this page also serve as a regression baseline: the same input must give the same result.
1. What is measured
Two things are looked at for every path.
- Candidate detection – did a path pattern find a potentially dangerous path in that account? This is the first classification, made from the permission combination alone, before any policy is applied.
- Verdict – the account's real policies (identity policies, permissions boundaries, SCPs, conditions) are applied to the detected path in the order of the AWS policy evaluation logic, and one of the four path verdict grades is assigned.
Candidate detection asks "is there a suspicious path?"; the verdict asks "does that path work right now?". The tables show the two separately.
2. Inputs and how fixtures are built
Two public sources are used.
| Source | Content | How it is used |
|---|---|---|
| pathfinding.cloud | A public catalog of AWS privilege-escalation paths (92 paths) | One synthetic account per path. |
| IAM Vulnerable | An intentionally vulnerable IAM environment (Terraform) | The privilege-escalation module is read as one account. 31 paths. |
2.1. One synthetic account per path
For each pathfinding.cloud path, one imaginary account is built.
- Required permissions – one user gets an inline policy that allows every permission the path requires (
permissions.required). If the path states a resource constraint, the permission is allowed on one ARN that fits it; otherwise on*. - Minimal targets – the targets a path needs (the role to pass, the target user, an administrator policy, and so on) are created only as far as needed, according to the path's category. No other resources are placed in the account.
- A fixed account ID is used.
2.2. IAM Vulnerable is read statically
The Terraform of IAM Vulnerable is never executed, only read as files. Users, roles, policies, policy attachments and groups are converted into the IAM data of one account, and variables and account-ID references are replaced with fixed values. The 31 escalation paths the environment documents are the expected list.
2.3. Three blocked variants
For every path with a detected candidate, a variant of the same account is built with one rule that blocks the path, to see whether the verdict really becomes Blocked.
| Variant | Rule added | Expected blocking layer |
|---|---|---|
| Account-level SCP deny | An SCP attached to the account denies the first required action of the path | SCP |
| Permissions boundary without the action | The principal's permissions boundary does not allow that action | Permissions boundary |
| MFA condition on the identity allow | The identity policy's allow requires aws:MultiFactorAuthPresent to be true (a leaked access key calls without MFA, so it is stopped) | Identity policy |
3. Results
3.1. Public escalation paths (one synthetic account per path)
The candidate detection (pattern) column shows how well a path pattern matched. Exact means the pattern's permission set equals the path's required permissions (read-only actions excluded), partial means the pattern's permissions are a subset, and none means no pattern applies. The three right-hand columns are the number of paths that became Blocked under each variant.
| Category | Paths | Candidate: exact / partial / none | Blocked under SCP deny | Blocked under boundary without the action | Blocked under MFA condition |
|---|---|---|---|---|---|
| new-passrole | 47 | 25 / 1 / 21 | 26/26 (100%) | 26/26 (100%) | 26/26 (100%) |
| existing-passrole | 23 | 16 / 0 / 7 | 16/16 (100%) | 16/16 (100%) | 16/16 (100%) |
| self-escalation | 8 | 8 / 0 / 0 | 8/8 (100%) | 8/8 (100%) | 8/8 (100%) |
| principal-access | 14 | 5 / 9 / 0 | 14/14 (100%) | 14/14 (100%) | 14/14 (100%) |
| All | 92 | 54 / 10 / 28 | 64/64 (100%) | 64/64 (100%) | 64/64 (100%) |
3.2. Grades of the unmodified account
Organization state measured means the organization was read and no SCP is attached; unmeasured means the guardrail policies were not read.
| Category | Organization | Decided | Decided (SCP not checked) | Candidate | Blocked | No finding |
|---|---|---|---|---|---|---|
| new-passrole | measured | 0 | 0 | 26 | 0 | 21 |
| new-passrole | unmeasured | 0 | 0 | 26 | 0 | 21 |
| existing-passrole | measured | 0 | 0 | 16 | 0 | 7 |
| existing-passrole | unmeasured | 0 | 0 | 16 | 0 | 7 |
| self-escalation | measured | 8 | 0 | 0 | 0 | 0 |
| self-escalation | unmeasured | 0 | 8 | 0 | 0 | 0 |
| principal-access | measured | 14 | 0 | 0 | 0 | 0 |
| principal-access | unmeasured | 0 | 14 | 0 | 0 | 0 |
| All | measured | 22 | 0 | 42 | 0 | 28 |
| All | unmeasured | 0 | 22 | 42 | 0 | 28 |
3.3. IAM Vulnerable (one account, escalation paths of the README table)
| Paths | Candidate detected | Decided | Decided (SCP not checked) | Candidate | Blocked | No finding |
|---|---|---|---|---|---|---|
| 31 | 31 | 7 | 0 | 24 | 0 | 0 |
4. Reading the results
4.1. What 100% under the blocked variants means
Under all three variants, the 64 paths with a detected candidate all became Blocked, and the verdict pointed at the layer of the added rule (SCP, permissions boundary, identity policy) as the place that stops them. In other words, when the account contains a rule that blocks, the engine does not miss it and leave the path as Decided or Candidate. This confirms that on this input there was no error in the dangerous direction (classifying a risky path as safe), which matches the grade rule that a "safe" verdict always rests on a confirmed deny.
4.2. Why the unmodified account has many Candidate paths
PassRole families (new-passrole, existing-passrole) cannot be settled by authorization rules, because the step where a service runs code with the role is outside them, so they are at most Candidate. Self-escalation and principal-access become Decided when nothing blocks them, and Decided (SCP not checked) when the organization's SCPs were not read. The difference between organization state measured and unmeasured shows in exactly these two rows.
4.3. Limits of candidate detection
Of the 92 paths, 28 have no pattern and are not detected, and 10 match only partially. These numbers show the coverage of the path pattern list, not of the verdict engine. The chained-AssumeRole path of IAM Vulnerable is detected by the role-chain pattern (2 to 3 hops), and every hop is judged against the trust policy and the calling principal's policies.
5. What this benchmark does not claim
- It is not a comparison with other tools. The tables show only input-to-result correspondence of this feature. No claim is made that it is better than, or finds more than, any other tool.
- Service-execution paths stay candidates. The step where a service runs code is outside the authorization rules, so these paths do not rise to
Decided. This is why 42 paths in the tables areCandidate. - Role chains are judged on the shortest chain only. Other chains that reach the same admin role are not listed, and chains longer than 3 hops are not detected.
- Expected grades are not ground-truth labels. Stored grades pin what this feature returned when the fixtures were generated, as a regression baseline. Correctness is checked on the blocked variants, where the existence of a blocking rule is known by construction.
- The results are for synthetic accounts and a vulnerable environment and do not represent the risk level of any real customer account.
6. Sources and licenses
- pathfinding.cloud – github.com/DataDog/pathfinding.cloud, Apache License 2.0
- IAM Vulnerable – github.com/BishopFox/iam-vulnerable, MIT License
- Basis of the evaluation rules – AWS policy evaluation logic
Both sources are used as read-only data. This page is neither official documentation of, nor endorsed by, those projects.