Back to AutoLLM
    Security Research

    Research & Whitepapers

    Original attack taxonomies, defense frameworks, and security advisories from AutoLLM's research team - published to advance the state of LLM security.

    Adversarial Red-TeamingRESEARCH

    ChaosProbeGPT: A Red-Teaming LLM That Attacks Like an Adversary

    Your AI should be broken in a safe room by us - long before a real crowd of attackers finds the same gap. A purpose-trained LLM that chains turns - reframing, pretexting - against your LLM or agent inside a controlled sandbox.

    4
    Attack Objectives
    <10ms
    Defense Overhead
    India
    Hosted Sandbox
    18 Nov 2026 3 min read
    Read paper
    ResearchRESEARCH

    How Do You Test Against Data You Cannot Touch?

    On-premise security keeps client conversations private. Here is the independent profile of the synthetic held-out set we built to test what our models actually understand.

    12,645
    Held-out Rows
    5,969
    Malicious Rows
    3
    Languages
    14 Sep 2026 7 min read
    Read paper
    Adversarial Red-TeamingCRITICAL

    Prompt Leak Arena: 300 Attacks, 9 Rewrites, All Broken

    GPT-4.1-mini Guardian vs Claude Sonnet 4.6 Mastermind. 300 attempts, 9 defense rewrites via APO - every version eventually broken. 10 full leaks. 45 partial leaks. The Guardian never held.

    300
    Attacks
    10
    Full Leaks
    9
    Rewrites Broken
    12 Apr 2026 15 min read
    Read paper
    Security AdvisoryHIGH

    "I Can't Share My System Instructions" Is Not a Security Control

    A structured refusal tells attackers everything they need to begin targeted extraction. Across 41 edtech deployments, system prompt recovery was achieved against 74% of them.

    74%
    Recovery Rate
    41
    Deployments Scanned
    8
    Median Attempts
    7 April 2026 5 min read
    Read paper
    Responsible DisclosureCRITICAL

    Soft Follow-up Exfiltration

    A novel prompt extraction attack exploiting social compliance and conversational stealth - achieving an 88% average vulnerability rate across 6 models with zero defenses triggered.

    88%
    Avg Vulnerability
    6
    Models Tested
    0
    Defenses Triggered
    19 Mar 2026 12 min read
    Read paper
    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.