According to Endor Labs, open source is playing a growing role across the AI tech stack, but most (52%) projects reference known vulnerable dependencies in their manifest files.
Latest from security vendors State of dependency management Just five months after its release, ChatGPT’s API is used by 900 npm and PyPI packages across “various problem areas,” according to reports, with 70% of these new packages being used.
However, just like any open source project, Endor Labs warns that the security risks associated with vulnerable dependencies must be managed.
“It’s great to see the latest generation of artificial intelligence APIs and platforms capturing the public’s imagination as they have in the past,” said Michael Sampson, principal analyst at Osterman Research. ‘s report is a clear demonstration that security has not caught up.”
“Advancement of technologies that identify potential weaknesses more quickly and automatically remediate them will make a big difference in this important area.”
For more information on malicious open source packages, see hundreds of malicious packages found in the npm registry.
Unfortunately, organizations appear to be underestimating the risks of not only AI APIs in open source dependencies, but security-sensitive APIs in general.
More than half (55%) of applications make calls to security-sensitive APIs within their code base, but the report claims that number rises to 95% when dependencies are included.
Endor Labs also warned that Large Language Model (LLM) technologies like ChatGPT are bad at scoring suspicious code snippets for malware potential. The results show that OpenAI GPT 3.5 has an accuracy rate of only 3.4%, while Vertex AI text-bison performs slightly better at 7.9%.
“Both models resulted in a significant number of false positives, which required manual review efforts and prevented automatic notification to each package repository to trigger package removal.” That said, the model appears to be improving,” the report said.
“These findings illustrate the difficulty of using LLM in security-sensitive use cases. Even if it goes up to 99%, it’s not enough to enable autonomous decision-making.”
Elsewhere in the report, developers point out that developers may be wasting time fixing vulnerabilities in code that isn’t even used by applications.
It claims that while 71% of typical Java application code comes from open source components, apps use only 12% of imported code.
“Vulnerabilities in unused code are rarely exploited. Organizations can eliminate or prioritize remediation efforts up to 60% by gaining reliable insight into which code can be reached across applications.” We can lower it,” the report said.
