AI agents still lie and cheat: new reward-hacking research warns smarter models aren’t more trustworthy – en.softonic.com