Researchers detail "context bombing", where defenders use prompt injections to trigger guardrails of attackers' LLMs, cutting AI hacking success rates by ~90% (Dan Goodin/Ars Technica)
·
0 reactions
·
0 comments
·
9 views
Dan Goodin / Ars Technica : Researchers detail “context bombing”, where defenders use prompt injections to trigger guardrails of attackers' LLMs, cutting AI hacking success rates by ~90% — Prompt injections, the malicious commands attackers embed into content to entice large language models to follow them …
Original article
Techmeme
Anonymous · no account needed
Discussion
0 commentsMore from Techmeme
Google agrees to buy 100% of the initial output from the Steel River Energy Center's 1.6GW solar project in Arkansas when it becomes operational in 2029 (Martha Muir/Financial Times)
Techmeme·1
eMarketer: chatbots like ChatGPT and Google AI Mode will generate less than $1B in ad revenue in 2026; OpenAI projected $2.5B in ChatGPT ad revenue this year (Trishla Ostwal/Adweek)
Techmeme·2
Sources: Nous Research, the startup behind Hermes, an open-source agent and OpenClaw rival, is raising $75M+ at a $1.5B valuation led by Robot Ventures (TechCrunch)
Techmeme·11
Q&A with Xinzhou Wu, head of automotive at Nvidia, on Nvidia's chips and AI models for autonomous driving, lidar's usefulness for Level 4 autonomy, and more (Nilay Patel/The Verge)
Techmeme·9
A look back at the 2007 heist of a Verizon data center in London, highlighting the role of physical data security as cyber defenses grow more sophisticated (Nathaniel Rich/New York Times)
Techmeme·10
How a Denmark-based research team is using quantum computers to accelerate AI protein discovery, showing a near-term commercial application for the quantum tech (Isabella Ward/Wired)
Techmeme·11