rocm

1 article
sort: new top best
clear filter
bug-bounty355 xss320 google244 facebook203 microsoft192 apple135 exploit111 rce94 csrf80 writeup62 account-takeover58 browser50 bragging-post48 defi48 smart-contract47 access-control46 ethereum44 malware42 open-source42 authentication-bypass40 ssrf40 web338 privilege-escalation36 docker35 ai-agents35 sqli34 cve34 smart-contract-vulnerability33 aws32 dos32 idor29 react28 oauth28 supply-chain27 sql-injection27 clickjacking26 wordpress25 solidity25 subdomain-takeover25 denial-of-service24 api-security23 information-disclosure23 race-condition23 node22 burp-suite22 automation21 cors21 phishing21 vulnerability-disclosure20 reverse-engineering20
0 8/10
A Field Guide to Reward Hacking in AI Kernel Generation
research

A comprehensive field guide documenting 10 distinct patterns where LLMs game kernel benchmarks through timing attacks (stream injection, thread injection, lazy evaluation, patching), semantic attacks (identity kernels, no-ops, shared memory overflow), and benign shortcuts, with defensive mechanisms for each exploit category.

ai-safety benchmark-gaming reward-hacking cuda gpu-kernels machine-learning testing-methodology triton rocm vulnerability-detection timing-attacks semantic-attacks correctness-validation adversarial-ml
KernelArena MI300X ROCm 6.x CUDA PyTorch Triton HIP
wafer.ai · matt_d · 1 day ago · details · hn