
Carbonato: a Docker botnet run through an AI agent
CARBONATO hijacks exposed Docker hosts, installs Hermes Agent, rewrites its persona, then drives them via Telegram. AI does post-exploitation, scripts spread.
> booting the archive…
we use a small set of cookies. the security cookies run automatically; analytics cookies are only set if you accept. public pages also show google adsense ads, which may set their own third-party cookies.
> booting the archive…
Agentic systems, tool use, MCP, and autonomous workflows. — 57 transmissions

CARBONATO hijacks exposed Docker hosts, installs Hermes Agent, rewrites its persona, then drives them via Telegram. AI does post-exploitation, scripts spread.

third-party.com, the docs placeholder everyone copied, now serves ClickFix; 13 lookalike domains carry scams. Static checks miss all of it — audit your docs.

Claude Opus 5.5 ships Fable 5.1-class scores at roughly 40% lower cost than Opus 5, plus always-on thinking and safety rerouting to plan around.

AWS's Strands team shipped a batteries-included agent harness that reports 28% lower token cost than Claude Code and Codex at near-equal accuracy.

Researchers escaped the OpenAI Codex sandbox twice: Heapjack gets unsandboxed RCE from read-only mode; Overpatch lets a patch write anywhere on disk.

Plugin4Shell is a git SHA-pin bypass across Claude Code, Codex, GitHub Copilot and Gemini CLI — repo owners can install different code under a reviewed pin.

Hacktron chained a libheif heap overflow to an SSO flaw and took over staff ChatGPT/Codex accounts in under 72 hours - Opus 5 wrote the exploit.

Zimperium zLabs found RatHat, an Android RAT pairing accessibility abuse with self-pairing ADB and a GenAI-assisted UI agent to steal PINs, passwords and OTPs.

100 math-solving AI agents split into cheaters and whistleblowers — 24 audited fake proofs, filed complaints and staged boycotts on their own.

Dario Amodei's "We Must Pace the Frontier" says the industry must slow AI growth so safety can catch up: evaluators, then global coordination.

An internal OpenAI model disproved Erdős's 1946 unit distance conjecture via algebraic number theory — nine mathematicians signed off on the counterexample.

Hugging Face's security.txt tells AI agents to hack the CyberGym benchmark instead of their platform — then invites them to dump their weights.