Flame malware used a chosen-prefix MD5 collision (2012)
The state-sponsored Flame cyber-espionage tool forged a Microsoft Windows code-signing certificate by constructing a chosen-prefix MD5 collision with a legitimate Microsoft Terminal Services certificate. Marc Stevens reverse-engineered the attack and published the cryptanalysis later that year.
SHAttered: two PDFs with the same SHA-1 (2017)
Stevens et al. at Google & CWI produced two PDF documents with identical SHA-1 hashes but completely different visible content. Total cost: about 6,500 CPU-years and 100 GPU-years of compute. Practical chosen-prefix SHA-1 collisions arrived three years later (Shambles, 2020) at a fraction of that cost.
Flickr 2009: signed URLs broken by length-extension
Flickr’s API used md5(secret || params) for API request signatures. Thai Duong and Juliano Rizzo showed that length-extension let anyone with one valid URL forge URLs with extra parameters , without knowing the secret. AWS S3 had a related vulnerability in their query-string authentication.
LinkedIn 2012: unsalted SHA-1 password leak
6.5 million LinkedIn passwords hashed with raw SHA-1, no salt, posted publicly. Within hours, ~90% had been cracked by rainbow tables. The incident is the canonical example used in every password-storage talk since.
Sony PlayStation 3: hardcoded random number
Not exactly a hash attack, but adjacent: Sony's ECDSA signature implementation reused the same nonce for every PS3 firmware signature. fail0verflow recovered Sony’s private signing key by spotting this in 2010, and the same construction made counterfeit signed firmwares trivial to produce.
Apple's NeuralHash was broken in days (2021)
Apple’s perceptual-hash design for client-side CSAM detection was reverse-engineered within hours of disclosure; collisions and preimages followed within days. Apple later shelved the client-side scanning plan, citing different concerns, but the episode is a textbook lesson that perceptual hashes are not cryptographic.