Researcher Pits 100 Abliterated AI Agents Against His Own Accounts in 5-Hour Hack Test
Summary
- • ~100 abliterated AI agents attempted to hack researcher's accounts over 5 hours
- • 5 accounts compromised: 3 via software vulnerabilities (IDOR), 2 via brute-force password attacks
- • Abliteration removes AI safety guardrails by subtracting the refusal vector from model weights
- • No tier-0 accounts (Gmail, banking) breached; no third-party zero-days discovered
Details
~100 agents, 5 hours; 5 accounts compromised, 16 social engineering attempts
Approximately 100 self-hosted AI agents ran for 5 hours; 5 accounts were compromised (3 via software vulnerabilities, 2 via brute-force), and 16 social engineering attempts were made.
Abliteration: removes AI refusal capability by subtracting a learned vector
Abliteration subtracts a 'refusal vector' — the model's learned rejection behavior — from its weights, removing safety guardrails while minimally impacting model intelligence, enabling full-capability hacking agents.
Three accounts breached via IDOR vulnerabilities in old side-projects
Agents identified and exploited IDOR-style vulnerabilities in old side-projects and hackathon demos, demonstrating that unmaintained legacy code remains a primary attack surface for autonomous agents.
Two accounts compromised via brute-force password attacks
Agents successfully brute-forced access to two accounts, demonstrating persistent password-based vulnerabilities as a viable attack vector even for unsophisticated autonomous agents.
Models: abliterated GLM-5.3, GLM-5.3 Flash, DeepSeek V4 Flash via Codex CLI
Three openly available abliterated models (GLM-5.3, GLM-5.3 Flash, DeepSeek V4 Flash from HuggingFace) were self-hosted on cloud GPUs, run via Codex CLI in individual Docker containers with provided cloud infrastructure keys.
No tier-0 accounts (Gmail, banking) breached; no zero-days discovered
Bounded results reflect current open-source model limitations: no tier-0 accounts (Gmail, 1Password, banking) were compromised, and no third-party zero-days were discovered — the experiment simulated an unsophisticated threat actor.
Researcher projects risks will grow as models improve and GPU costs fall
Results reflect today's open-source model capabilities; the researcher expects risks to increase substantially as models improve and GPU costs fall, making such attacks more accessible, effective, and likely to breach higher-value accounts.
AI agent red-team experiment — sourced from TLDR AI / Shrivu Shankar blog post, September 2026.
What This Means
This experiment provides a concrete, reproducible data point on the current offensive capability of unconstrained self-hosted AI agents: with freely available abliterated open-source models, a few hundred dollars of cloud GPU time, and only a target's name, an unsophisticated actor can compromise multiple accounts within hours. The success against old, unmaintained code is particularly instructive — most people have legacy projects and demos with real vulnerabilities that autonomous agents can now surface and exploit at scale. The trajectory is one-directional: as open-source model quality improves and GPU costs continue to fall, these attacks will become cheaper, faster, and more effective — likely crossing into higher-value account territory within a few model generations.