#backdoor

Adversarial Training Can Further Hide a Backdoor, but Does It Improve Trigger Identification?

Summary The Sleeper Agents paper demonstrated that adversarial training on red-team prompts failed to remove a planted backdoor. The authors further suggested that it taught the model to "better identify when to act unsafely." This would pose a significant danger, as a standard safety practice would be refining the very behavior it set out to remove. However, the degree of danger depends on exactly what the model is learning during adversarial training. We discuss how, due to saturation of firing rates, more benign scenarios such as the model's backdoor response being generally weakened are al

LessWrong · EN

Botnet XCPE - La backdoor restera en place dans les box Android TV

Google et ses partenaires ont coupé les serveurs du botnet XCPE et banni Videostrong, qui avait piégé le firmware de plus de 1,3 million d'appareils Android. Mais l'implant reste dans les box déjà vendues, et les partenaires ne donnent pas la même parade : reset d'usine pour Plume, poubelle pour Synthient.

Korben · FR

cPGuard Update Channel Hijacked to Backdoor Linux Servers and Collect Password Hashes

An attacker broke into the update system of the cPGuard security tool on November 2 and used it to send a malicious script to customers' Linux servers. The script installed a backdoor and collected password hashes and other sensitive files from servers that ran it, cPGuard maker OPSSHIELD said in an advisory. The attacker has since taken some of that data. cPGuard 5.88.25 and later versions

The Hacker News · EN