AI Risks
Security vulnerabilities identified by Claude Mythos
This is not a drill. Since giving my latest talk - AI in Telecommunications: The New Attack Surface - it's been revealed that #Mythos, the latest model from #Anthropic, has been able to uncover numerous security issues in software that goes back over 20-30 years. In fact it ‘saturates’ all existing methods of testing how good a model is at offensive cyber capabilities. That is to say it scores close to 100%, so these tests no longer determine how effective it is.
The question is now how far will it go to create security exploits? I discussed concepts such as prompt injection, data poisoning and reworking the actual models to insert exploits into code that is generated with model manipulation.
Time will tell , but as recommended we must take a zero trust approach to any #AI code generation and test it thoroughly.
In the mean time - stay frosty
