Breaking

AI Jailbreak Researcher Pliny Releases Multi-Agent Offensive-Security Framework T3MP3ST

July 5, 2026 at 06:06 EDT

Security
AI Agents
Open Source

Pliny (elder_plinius), known as an AI jailbreak researcher, on July 5, 2026 released T3MP3ST on GitHub, a multi-agent offensive-security framework that repurposes an existing Claude Code or OpenAI Codex coding-agent subscription as its "brain" to autonomously run reconnaissance, exploitation and reporting. It is offered as fully free, open source under the AGPL-3.0 license.

July 5, 2026 · Pliny the Liberator

T3MP3ST turns your AI coding agent into an autonomous bug-hunter

A free, open-source multi-agent offensive-security framework that runs the full kill chain — recon to exploit to report — using a Claude Code, Codex, or Hermes subscription you already own as the "brain." No extra API keys. Authorized use only.

specialist operators in the "swarm" — recon, scan, exploit, lateral, exfil, persistence, C2, report

35→83

default tools, expandable to the full arsenal (nmap, nuclei, semgrep, ffuf…)

AGPL-3.0, self-hosted, "not pay-to-play" — no additional API keys required

CVE-Zero · 10 new CVEs disclosed after training cutoff, untuned

Single agent vs. swarm — CVEs identified to file / line / CWE

8 / 10

Single agent

10 / 10

Swarm of 8

Across 7 languages · web apps/APIs, network recon, source-code audit, CTF, smart contracts, embedded/IoT/OT/SCADA.

Benchmark results · pass@1 · reproducible via verify-claims

XBEN white-box (min 102/104)98.7%

XBEN black-box (beats XBOW's own 85%)90.1%

Cybench (Opus 4.8, hints removed)58%

The kill chain, orchestrated in natural language

Op Admiral (plans) → Shared blackboard → Recon → Exploit → Report → War Room (Web UI)

Follows the Cyber Kill Chain & MITRE ATT&CK · CLI + HTTP API + MCP · installs via npm · high-risk tools (Metasploit, Hydra) gated behind human approval.

Why it matters

Developers call it a major boon for solo founders and small teams — zero-day hunting with an existing subscription, no costly specialist tooling.

The caveat

Pliny notes the swarm's end-to-end reliability is still unbenchmarked and experimental; coordinated swarm and some features remain in development in v1.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.