Tag: #llm-security18 results
A developer seeks recommendations for high-performing LLMs to generate adversarial prompts and requests public 'golden' datasets for benchmarking AI agent security. The inquiry focuses on automating red-teaming for vulnerabilities like prompt injection, jailbreaks, and tool misuse.
This article explores how security teams can leverage Amazon Bedrock to combat the rising threat of sophisticated, AI-generated phishing emails. It highlights the integration of generative AI and OSINT to identify malicious patterns.
OpenAI has introduced safety updates to ChatGPT designed to better recognize context during sensitive interactions. These improvements help the model detect potential risks over time and provide safer, more reliable responses.
Cloudflare is opening its advanced Client-Side Security tools to all users. The new cascading AI detection system combines graph neural networks and LLMs, reducing false positives by up to 200x. It excels at catching sophisticated zero-day exploits.
China's 315 gala exposed GEO scams where 10+ fake articles generated by software fool LLMs into recommending fictional products like AstroTekk Apollo-9 smartband. Costs start at 299 RMB; AI treats repeated fakes as consensus. Highlights risks in product recommendations and competitor smears.
This research-oriented post offers a mechanistic explanation of prompt injection and argues that understanding model roles is essential to studying the vulnerability. It encourages practitioners to examine how models distinguish system, developer, user, and other instruction sources.
Post queries if privacy-preserving AI/ML demand has risen with LLMs, regulations, and de-anonymization risks. References old discussion and anecdotal use of trusted execution environments for enterprise LLMs. Seeks subreddit opinions on AI privacy trends.
Proposes authenticated prompts and context for cryptographic provenance in LLM apps. Features policy algebra with Byzantine resistance and layered defenses. Achieves 100% attack detection with zero false positives.