Ptolemy

Offensive AI Security

We break AI agents. And everything they touch.

Offensive security across web apps, APIs, and the AI agents wired into them. We find what scanners miss and most teams never look for, then write down exactly what gave way, in a form you can hand to an auditor, a customer, or a regulator.

  1. The Surface

    Three surfaces. One attacker.

    Web. APIs. AI agents. Security testing was built for the first two. Now agents wire all three together, and the gaps between them are where things break. We test the full surface, not the pieces.

  2. The Chain

    A misbehaving agent is a finding. An agent that opens a door is a breach.

    We don't stop at "the agent said something wrong." We follow the compromise into the systems it connects to, mail, CRM, tickets, databases, and show you the full path from entry to impact.

  3. The Report

    Every finding is reproduced, not inferred.

    Numbered steps. Replayable attacks. Severity grounded in what the path actually reached. Mapped to the EU AI Act, NIST AI RMF, and OWASP LLM Top 10. Then we re-test after you fix it.

  4. First Light

    Everything hidden comes to light eventually.

    The only question is who finds it first. Don't take our word for it; our research is published, our findings are credited by Apple and Google security programs, and our only product is the truth about your system.

Models that leaked customer data
7 of 8
Adversarial test executions published
5,264+
Published studies
3

AI agents are being wired into the systems Western commerce depends on: healthcare, finance, insurance, logistics.

The companies building those agents will tell you they are safe. Our research says otherwise. We exist to close that gap.

In Practice

What an unprotected agent does when asked nicely.

Three attacks, two outcomes. Toggle between the agent as deployed and the same agent after assessment and remediation.

Support assistant

Web chat

Simulated exchanges based on attack patterns from real assessments.

See where an attack goes after the reply →

What We Test

If an agent can read data or take an action, it can be attacked. Which one is yours?

Support and sales assistants that can see account data and act on it.

See what we test for it →

We publish what we find.

Attacker-first

Every assessment starts from the outside. We map what an attacker can reach before we read a single architecture doc, because the threat doesn't get a briefing.

We find the problem, not the fix

We don't sell you the remedy for what we find. The report has no conflicts because we have none. The people who build a system shouldn't be the only ones grading it.

The methodology is public

Three published studies. 5,264+ adversarial test executions. Findings credited by Apple and Google security programs. The methodology is public because it holds up.

5,264+

Adversarial test executions

3

Published studies

7 / 8

Models leaked customer data

They fixed vulnerabilities with us

  • Comex Logistics
  • Immaculate Shipping Line

Research

October 2026

Absent Deliberation

How LLMs Resolve Competing Objectives Without Reasoning About Them

LLM customer service agents confirm stored PII when asked to verify it, while refusing direct requests for the same data. We trace the mechanism through more than 1,500 API calls across six Claude models with full thinking-trace capture. The model almost never asks itself whether confirmation is disclosure, and when it does, it reasons past the question. Made to take a position on the scenario beforehand, it refused in every trial. More reasoning effort did not reliably help, and in one model made disclosure more explicit.

Read →

September 2026

Confirmation as Disclosure

Identity-Assumption Attacks Against LLM Customer Service Agents

An attacker states customer data and asks the agent to confirm it, and the model verifies it without authenticating the requester. We evaluate this vulnerability class across eight models from two providers with over 1,160 test executions. Seven of eight models disclosed PII in at least one reported condition, and a configuration adapted from a published customer-support guide disclosed in 84% of tests. Verification instructions in the system prompt did not prevent it; application-layer authorization did.

Read →

April 2026

Red Team Evaluation of an AI Agent Security Layer

Across Regulated Industries

Findings from a structured adversarial evaluation of an AI agent security layer across four regulated industries, with over 2,600 attack payloads executed across multiple agent configurations. The guard achieved a 90.6% combined defense rate with a 0% false positive rate on benign traffic.

Read →