Search AI company Elastic Security Labs has released a report on avoiding large language model (LLM) risks and abuses. The LLM Safety Assessment: The Definitive Guide on Avoiding Risk and Abuses offers attack mitigation best practices and suggested countermeasures for LLM abuses.
Elastic says generative artificial intelligence (AI) and LLM implementations have become widely adopted over the past 18 months, with some companies pushing to implement them as quickly as possible. This has expanded the attack surface and left developers and security teams without clear guidance on adopting emerging LLM technology safely.
“For all their potential, broad LLM adoption has been met with unease by enterprise leaders, seen as yet another doorway for malicious actors to gain access to private information or a foothold in their IT ecosystems, said Elastic’s Jake King. “Publishing open detection engineering content is in Elastic’s DNA. Security knowledge should be for everyone. We hope that all organisations, whether Elastic customers or not, can take advantage of these new rules and guidance.”
Elastic’s search, observability and security solutions use the Elastic Search AI Platform, a development platform used by thousands of companies, including more than 50% of the Fortune 500.
The LLM Safety Assessment builds and expands on the Open Web Application Security Project (OWASP) research focused on the most common LLM attack techniques. The research includes crucial information security teams can use to protect their LLM implementations, including in-depth explanations of risks, best practices and suggested countermeasures to mitigate attacks.
The countermeasures explored in the research cover different areas of the enterprise architecture, primarily in-product controls, that developers should adopt when building LLM-enabled applications and information security measures that SOCs must add to verify and validate the secure usage of LLMs.
In addition to 1000+ detection rules already published and maintained on GitHub, Elastic Security Labs added an initial set of detections just for LLM abuses. These new rules are an example of the out-of-box detection rules now included to detect LLM abuses.
You can read the full report here.
