Banks all over the world can be hacked and PIN numbers seized by exploiting a flaw in a common cryptoprocessor made by IBM, researchers from Cambridge University have found.

The cryptoprocessor in question, the IBM 4758, sits at the end of cash machines and scrambles the PIN number that people type in, as well as the program used to verify the PIN at the other end.

It uses a minimum 64-bit encryption (112-bit for keys), conforms to the US Data Encryption Standard FIPS 140-1 Level 4 and was thought to be impossible to crack, even physically. But the problem does not lie with the cryptoprocessor but the software it runs on - Common Cryptographic Architecture (CCA), which comes free with the 4758.

The CCA software requires two or more people to combine their access priviledges before any security changes can be made. As such, it was thought that codes could only be cracked if there was collusion between two high-level bank employees.

The Cambridge researchers - Michael Bond and Richard Clayton - found however that they could persuade the 4758 to send its encryption key (stored only within the cryptoprocessor itself) to them if it was itself encrypted with a 3DES (triple DES security) key.

All you need to know is the value of the "exporter key". Thus, with "a mixture of sleight-of-hand and raw processing power", you can get hold of a cryptoprocessor's key and from that point you can access to every piece of data sent through it - which could be hundreds of thousands of card numbers with their relevant PINs.

The only problem then remaining is getting a high-enough level of access to a bank's computer system. Not easy, but not impossible either.

Messrs Bond and Clayton reckon they need 20 minutes at a console to get the information and then two days or so to break it down. All you need is some nous and a standard $995 Altera FPGA evaluation board.

"A crooked bank manager could duplicate our work on a Monday and be off to Bermuda by Wednesday afternoon," Mr Clayton said.

The researchers posted full and extensive details of their methods on the Internet last night, stating that "our main reason for publishing this attack is to demonstrate the power of our parallel key search machine". They also allege that IBM was informed of the possibility of such an attack a year ago but has failed to fix the problem or produce any patches.

That now looks very likely to happen and banks that use the 4758 will have to upgrade security immediately at huge cost.