Google has announced that it is open-sourcing Magica, an artificial intelligence (AI)-powered tool for identifying file types, to help defenders accurately detect binary and text file types.
“Magika outperforms traditional methods of file identification, delivering up to 95% accuracy and boosting overall accuracy by 30% on traditionally difficult to identify, but potentially difficult content such as VBA, JavaScript, and Powershell. demonstrates,” the company said.
The software uses a “custom, highly optimized deep learning model” that enables accurate identification of file types in milliseconds. Magica implements inference functions using Open Neural Network Exchange (ONNX).
Google said it uses Magicka extensively internally to improve user security by routing Gmail, Drive, and Safe Browsing files to appropriate security and content policy scanners.
In November 2023, the tech giant unveiled RETVec (short for Flexible and Efficient Text Vectorizer), a multilingual text processing model in Gmail for detecting potentially harmful content such as spam and malicious emails. .
Amid ongoing debate over the dangers of the fast-developing technology and its misuse by nation-state actors linked to Russia, China, Iran and North Korea to promote hacking efforts, Google said AI should be widely used. Deploying on can strengthen digital security and “tilt” the cybersecurity balance from attackers to defenders.
He also emphasized the need for a balanced regulatory approach to the use and adoption of AI to avoid a future where attackers can innovate, but defenders are held back by AI governance choices.
“AI allows security professionals and defenders to measure their work in threat detection, malware analysis, threat detection, threat fixing and incident response,” Phil Venables of the tech giant and Royal Hansen noted. “AI offers a great opportunity to overcome the defenders’ dilemma and tilt the scales of cyberspace to give defenders a decisive advantage over attackers.”
Concerns have also been raised about the use of web scraped data by AI models developed for training purposes, which may include personal data.
“If you don’t know what your model is going to be used for, how can you ensure that its downstream uses will respect data protection and people’s rights and freedoms?” The UK Information Commissioner’s Office (ICO) pointed out last month.
What’s more, new research has shown that large language models can act as “sleeper agents” that may appear innocuous, but when certain criteria are met or special instructions are given, they can be activated. Can be programmed to engage in deceptive or malicious behavior.
“Such backdoor behavior can be made permanent so that it is not removed through standard safety training techniques, including supervised fine-tuning, reinforcement learning, and aversive training (eliciting the unsafe behavior and then removing it). training) is involved,” said researchers at AI startup Anthropic. in the study.