All Research
    Adversarial Red-Teaming · Attack Generation · Chaos Probe

    ChaosProbeGPT

    A red-teaming LLM that never gets tired and never runs out of ideas

    Your AI should be broken in a safe room by us - long before a real crowd of attackers finds the same gap.

    Published

    18 November 2026

    Published By

    AutoLLM Security Research

    Availability

    Enterprise only

    Sandbox

    Hosted in India

    01 · The Attacker

    Built the attacker first

    ChaosProbeGPT is a purpose-trained LLM that generates live attacks against an enterprise LLM or agent - inside a controlled sandbox, without the refusals a general-purpose model throws the moment the objective gets uncomfortable. You give it a target and a goal, and it works the goal.

    A red-teaming team works in shifts. It writes its report and moves on to the next engagement. ChaosProbeGPT stays on the target - it does not get tired, it does not stop at the first refusal, and it keeps looking for the next way in.

    A defense is only as strong as the attacks it has survived. So we built the attacker first - and every attack it lands informs the guardrails, perpetually. The attacker never stops, so neither do the defenses.

    02 · The Campaign

    You set the objective. It runs the campaign.

    Persuade the system to leak information it was told to protect
    Role-manipulation - talk the model out of its own instructions
    KYC bypass
    Authorize money movement where an agent has the ability to act

    It doesn't fire one payload and stop. It chains turns - reframing, pretexting, walking a model from a polite refusal to a fully populated sanction letter three messages later. That's how real compromise happens. That's what we test.

    ChaosProbeGPT generating a multi-turn attack plan against a lending assistant, with techniques, targeted boundaries and fallback messages per turn
    A working attack generation: ChaosProbeGPT plans a turn-wise campaign against a loan assistant - each turn with a technique, the boundary it targets and a fallback if refused.
    03 · In Production

    This isn't theoretical.

    In audits of live production fintech systems we've already watched AI agents generate approved sanction letters, and voice agents leak their internal discounting hierarchy straight out of the system prompt, mid-call. The discount-manipulation pattern - a chatbot conjuring concessions that exist in no pricing rule - reproduces across multiple Indian deployments. ChaosProbeGPT is how we surface these before someone else does.

    04 · Deployment

    Built to run where your workflows run.

    ChaosProbeGPT is a fine-tuned open-source model we host ourselves. No data leaves your environment, no dependence on a third-party API, usable against the private, mission-critical workflows you can't hand to anyone's cloud. The sandbox is hosted in India.

    05 · The Honest Part

    The honest part.

    No defense system in the world can't be breached. Anyone selling you breach-proof is either lying or doesn't know their own failures. Our attack library is indigenous, already past 2 lakh batteries, and growing - because this is a cat-and-mouse race, not a product you buy once. We publish our failures. Our defenses are deterministic and add under 10ms. You will see exactly where they hold and where they don't.

    06 · Launch

    Launching now - enterprise only.

    Free trial for BFSI and fintech teams willing to exchange findings with us under NDA. You run a mission-critical LLM or agent in production; we run ChaosProbeGPT against it in the sandbox; you get the findings, and every finding feeds back into the attack library and the guardrails. Responsible disclosure, both ways.

    If you're deploying AI where money moves or identity is verified, you want to meet your attacker in a sandbox - not in an incident report.

    07 · What's next

    A technical report is on the way.

    This note is the opening move. A full technical report follows soon - success rates from production deployments, the attack objectives that landed most often, and how the guardrails adapted after each one. Real numbers from real systems, shared with the same honesty as everything above.

    Follow this update

    Be the first to read the production report when it lands.

    AutoLLM · Chaos Probe · ChaosProbeGPT

    Book a ChaosProbeGPT trial

    Enterprise only. Free for BFSI and fintech teams exchanging findings under NDA.

    Published by AutoLLM Security Research · 18 November 2026
    TOI AI Quotient Awards 2026 NomineeFeatured in Times of India · AI Quotient Awards 2026 Nominee →

    Need Custom Fine-Tuning or On-Premises Deployment?

    Get enterprise-grade security with complete data control. Custom fine-tuning and on-premises deployment available for Enterprise customers.