LLM & agent red teaming
Prompt injection, indirect injection, tool misuse, multi-turn attacks, and authorization failures.
Test plan · transcripts · confirmed findingsAI security consulting
AI security testing for models and agents.
Cordon Security finds prompt-injection, unsafe tool use, and authorization failures in LLM applications and AI agents. For client work, Bill runs the test with ARES and reviews every finding.
ARES track recordResults counted once per tested model.
Distinct models tested25
Models with a confirmed break≥15
ARES development beganDec. 2025
Services
What Cordon Security tests.
Cordon Security works in its own lab or against systems you have authorized Bill to test. The focus is on AI systems that read untrusted data, call tools, change state, or make security decisions.
Exploit validation
Bill reproduces suspected failures in an isolated lab or an approved environment and checks them against a defined success condition.
Reproduction steps · impact · limitsRemediation & retesting
Turn confirmed findings into workflow changes and regression tests, then rerun the test to see whether the fix holds.
Fixes · regression tests · retestTrack record
At least 15 of 25 models produced a confirmed break.
Since December 2025, ARES-generated attacks have produced confirmed breaks on at least 15 of the 25 distinct models Bill tested.
Model red teamingARES-generatedDec. 2025 to present
ARES model-testing record
Each test had a defined success condition. A model counted only after the platform or test judge confirmed that condition. ARES generated the counted attacks, and Bill reviewed the results.
Tasks and attempt budgets varied, so this number reports coverage across models.
25Distinct models tested
≥15Models with a confirmed break
ARESGenerated every counted attack
BillReviewed the results
Process
How an assessment works.
Bill sets the boundaries and runs the engagement. ARES generates and adapts attacks inside that scope.
Scope
Agree on targets, success criteria, data rules, stop conditions, and who can approve changes.
Map
Identify tools, untrusted inputs, permission boundaries, and actions that can change state.
Test
Run approved attacks against fresh test state and within the agreed budget.
Confirm
Check the actual system action. Persuasive text alone does not count as a break.
Report
Document the finding, recommend a fix, and rerun the test after remediation.
Operating rules
How Bill runs the work.
Cordon Security is a security consulting practice of Compass Point Ventures LLC. ARES is an internal tool operated by Bill.
Written permission
Testing starts after the target, techniques, time window, and stop conditions are agreed in writing.
Bill controls execution
ARES generates attacks. Bill approves the work, reviews the evidence, and decides what enters the report.
Controlled data
Models and traces can run in infrastructure controlled by Cordon Security, with data handling set before testing begins.
Complete reporting
Reports include confirmed findings, failed tests, unresolved questions, and known limits.
Contact
Tell Bill what you want tested.
Include the type of system, what you are concerned about, and whether you own it or have permission to test it.