← Voltar ao feed
medium

Researchers show 'sleeper agent' technique hiding malware payloads inside open-weight AI model parameters

1 fonte(s) · 1 atualização(ões) · publicado 16/11/2026 16:10 · coletado há 23 horas

Baseado só no resumo do feed. O site da fonte não liberou o texto completo para o Tifon, então esta análise saiu apenas do resumo publicado no feed e pode estar incompleta. Ela é atualizada quando outra fonte cobrir o assunto. Ler na fonte original ↗

A Dark Reading report describes a 'sleeper agent' tactic for hiding malicious payloads inside open-weight AI models. Per the report, malware can occupy as little as 0.00001% of a neural network's parameters, which makes it very hard for users to detect. The article text supplied is only a headline and a one-line teaser. No web research was run, so there is no information on the researchers, the method, affected model hubs, CVEs, indicators or exploitation in the wild. The item reads as a research-style or proof-of-concept supply-chain risk, not a confirmed campaign. Defenders should treat models downloaded from public repositories as untrusted artifacts.

Remediação

The article text does not state any remediation. General practice is to source models only from trusted publishers, verify hashes and signatures, prefer safe serialization formats, and scan or sandbox models before loading them.

Linha do tempo

  1. darkreading.com ↗ · pub. 16/11/2026 16:10 · coletado 16/11 17:00 primeiro registro só resumo do feed

    Researchers show 'sleeper agent' technique hiding malware payloads inside open-weight AI model parameters

    ler a análise desta fonte

    A Dark Reading report describes a 'sleeper agent' tactic for hiding malicious payloads inside open-weight AI models. Per the report, malware can occupy as little as 0.00001% of a neural network's parameters, which makes it very hard for users to detect. The article text supplied is only a headline and a one-line teaser. No web research was run, so there is no information on the researchers, the method, affected model hubs, CVEs, indicators or exploitation in the wild. The item reads as a research-style or proof-of-concept supply-chain risk, not a confirmed campaign. Defenders should treat models downloaded from public repositories as untrusted artifacts.

    Remediation

    The article text does not state any remediation. General practice is to source models only from trusted publishers, verify hashes and signatures, prefer safe serialization formats, and scan or sandbox models before loading them.