I Asked 100 Agents to Hack Me (9 min read)

Shrivu Shankar conducted a 5-hour hacking experiment on his own accounts using approximately 100 autonomous AI agents based on abliterated open-source models. The experiment resulted in the compromise of 3 accounts via software vulnerabilities and 2 accounts via password brute forcing, along with successful cross-verification of sensitive personal information.

Experiment Setup and Methodology

The experiment utilized abliterated models from Hugging Face, such as GLM-5.3 and DeepSeek V4 Flash, with each agent running in an individual Docker container. The agents were connected to Codex CLI to perform internet searches, OSINT (Open Source Intelligence) gathering, and vulnerability scanning. Abliteration technology adjusts models to maintain intelligence and perform offensive tasks without safety measures by subtracting the 'refusal' vector from model weights.

Key Achievements and Limitations

The agents accessed accounts by discovering IDOR vulnerabilities in past hackathon projects and poorly managed credentials. They also collected personal information such as phone numbers and addresses by combining free databases like FastPeopleSearch with social media crawling. However, they failed to discover third-party zero-day vulnerabilities and did not compromise Tier 0 accounts with enhanced security, such as Gmail or 1Password.

Costs and Implications

The experiment cost approximately $210 for 5 hours of usage on 2xB300 GPUs, achieving successful access at a cost of about $40 per account. The author predicts that future optimization of smaller models could reduce costs to under $5 per account. This suggests that the lowering barrier to entry for open-source agents may increase the frequency of attacks by non-experts, emphasizing the need for defensive measures such as moderation guardrails and KYC regulations from inference providers.