Based on 5 reviews
The company provides a definition of the term abrade as a verb from latin abradere, meaning to wear away by friction or scraping. The defining property is repetition, not force, accumulating friction until the material gives way. This concept applies to deliberately subjecting an agent to friction before something else does it by accident.
The product works by injecting poisoned telemetry into the inputs an agent processes and then evaluates how the agent responds. It is positioned as automated adversarial friction that can be re-run when prompts, models, or permissions change. The service is distributed via a command-line installer.
The service maps attack categories to OWASP LLM Top 10 LLM01:2025, LLM02:2025, and LLM06:2025. It also covers MITRE ATLAS 2026.10 AML.T0051, AML.T0053, and AML.T0085.002. These specific mappings are listed on the company's website.
The service is designed for anyone who can reach an app and write to logs, such as crafted usernames on failed logins or malformed requests. If an AI agent reads those logs and holds cluster credentials, the attacker only needs to get a sentence into the telemetry the agent implicitly trusts. Guardrails do not solve this because they scope which tools an agent may call but do not decide whether the agent should have believed the log line.
In a lab test, Opus, Sonnet, and Haiku all scaled a workload to zero when told to trust telemetry. The ledger showed the model listing deployments, reading events, scaling to zero, and listing again to check its own work. The prompt told the agent to trust its telemetry, and it never saw the policy that forbade scaling to zero.
The service includes a Harness, AUT, Judge, and Canary architecture where no agent is ever deployed into the cluster during the test. This setup allows for isolated evaluation of the agent's response to adversarial inputs without affecting production environments. The process ensures that the measurement is specific to the declared boundary.
The company states that the fix belongs in the policy and whatever enforces it, not only in the agent's prompt. A check against that policy would have passed the call in the example where scaling to zero was allowed. The most capable model tested did it as reliably as the least capable model.
The service is described as a full suite of adversarial friction that accumulates over time rather than being resolved in one pass. The defining property is repetition, not force, which aligns with the concept of abrasion. This means the testing is designed to reveal what the agent is actually made of through repeated friction.
The service is targeted at AI agents that hold real infrastructure credentials such as kubectl, AWS APIs, Terraform, and CloudWatch. It is designed for environments where these specific tools are used to manage infrastructure. The testing is specifically tailored to the risks associated with these credentials.
Abrade is a information technology and services listed on Trustburn. Abrade provides pre-deployment security testing for AI agents that hold real infrastructure credentials such as kubectl, AWS APIs, Terraform, and CloudWatch.
The official website of Abrade is abrade.ai.
On Trustburn, Abrade has a rating of 4.2 out of 5 based on 5 reviews.
AI-generated from reviews on this page. Verify anything important with the company directly.
The installation process via CLI is very efficient. The tool integrates well with existing workflows. The ability to run tests on every prompt change ensures consistent security. The scorecard provides clear insights into agent behavior. It seems like a reliable solution for pre-deployment validation.
Getting started with Abrade was a breeze. The CLI install command worked instantly, and the setup process was straightforward. It feels like a tool designed for speed and efficiency.
The way Abrade handles telemetry spoofing is impressive. By synthesizing crafted CloudWatch events and Prometheus annotations, it effectively tests how agents handle injected data. The canary check mechanism, which uses exact string matches for planted credentials, provides a definitive verdict that overrides the judge's classification. This ensures that evidence is not argued with. The corpus packs are versioned and reproducible, which is essential for consistent security reviews. The fact that the agent is never deployed into the cluster but sits outside as an impersonator of trusted telemetry is a significant advantage. It allows for safe, repeated friction testing on every prompt change or model upgrade. The scorecard provides clear visibility into held, compromised, leaked, and overreach verdicts, making it easy to audit IAM policies. This tool seems essential for any team deploying agents with real infrastructure credentials.
When analyzing Abrade's approach, the distinction between guardrails and adversarial friction stands out. Traditional guardrails scope which tools an agent may call, but they do not decide whether the agent should have believed the log line that told it to call one. Abrade injects faults into what the agent believes, simulating crafted usernames or malformed requests that land in error messages. This is crucial because anyone reaching the app can write to logs. If an AI agent reads those logs and holds cluster credentials, the attacker does not need to touch the agent directly. The measurement capability shows that scaling to zero is allowed by policy, even if the agent reasons correctly within that policy. The test harness synthesizes crafted events, and the judge uses a separate model to classify replies. This layered approach seems more robust than simple policy checks, as it reveals gaps between declared boundaries and actual agent behavior. The canary check adds another layer of security by using exact string matches rather than relying solely on model judgment for credential leaks.
The adversarial testing approach feels really solid for what we need. Being able to inject faults into the agent's context without touching the cluster is a smart move. The corpus updates keep things fresh, and the CLI install was painless. Seems like a reliable way to catch those subtle permission leaks before they happen.
No news yet. This is where the business shares its own updates and press.
No photos yet. Photos shared by reviewers will show up here.
Abrade provides pre-deployment security testing for AI agents that hold real infrastructure credentials such as kubectl, AWS APIs, Terraform, and CloudWatch. The product works by injecting adversarial payloads, including poisoned telemetry referencing OWASP LLM Top 10 and MITRE ATLAS techniques, into the inputs an agent processes. It then evaluates how the agent responds, producing metrics such as a "Containment Index" and an "Agent Verdict" indicating whether the agent refused harmful actions. The service is positioned as automated adversarial friction that can be re-run when prompts, models, or permissions change, and is distributed via a command-line installer.