Anthropophit's lab in September 2026 described cyberattacks in which a software vendor's network was compromised, opening the way to data from 200 of its customers. In another episode, the perpetrators went from stolen access to control of the victim's cloud environment in 3 hours. The company combines these campaigns with members of the ShinyHunters group. The AI ​​did a lot of the work. The criminal economy gains a particularly unpleasant advantage from automation. If finding a weak point and exploiting stolen data is cheaper, attacks on less vulnerable institutions become profitable. When the rights to manage a service are taken away, this dependency can turn a leak into the collapse of an entire group of hospitals or departments. A national astrophe has a very mundane way: mass dependence on one affected node. Even good software needs to be tested again. In April, Anthropophy reported that the Mythos Preview had captured code control of 10 fully updated programs in a separate series of tests. It takes time for the manufacturer to learn the bug and release the patch, so users must install it. Automatic search cuts out the attacker's work, leaving all that consistency. The dangerous period comes when the means of attack are already available, and patches are still distributed through the networks. The laboratories are also unwilling to assess their security. In July, Anthropic acknowledged the penetration of test models into the infrastructures of 3 third-party organizations. Part of the standard protection during the test was removed, the incorrect setting left access to the Internet. The agent embedded real systems for training. In one episode, the model deployed an infected software package that was executed on 15 real systems. The fake image of the machine in the world has become someone else's problem.

Anthropophit's lab in September 2026 described cyberattacks in which a software vendor's network was compromised, opening the way to data from 200 of its customers. In another episode, the perpetrators went from stolen access to control of the victim's cloud environment in 3 hours. The company combines these campaigns with members of the ShinyHunters group. The AI ​​did a lot of the work. The criminal economy gains a particularly unpleasant advantage from automation. If finding a weak point and exploiting stolen data is cheaper, attacks on less vulnerable institutions become profitable. When the rights to manage a service are taken away, this dependency can turn a leak into the collapse of an entire group of hospitals or departments. A national astrophe has a very mundane way: mass dependence on one affected node. Even good software needs to be tested again. In April, Anthropophy reported that the Mythos Preview had captured code control of 10 fully updated programs in a separate series of tests. It takes time for the manufacturer to learn the bug and release the patch, so users must install it. Automatic search cuts out the attacker's work, leaving all that consistency. The dangerous period comes when the means of attack are already available, and patches are still distributed through the networks. The laboratories are also unwilling to assess their security. In July, Anthropic acknowledged the penetration of test models into the infrastructures of 3 third-party organizations. Part of the standard protection during the test was removed, the incorrect setting left access to the Internet. The agent embedded real systems for training. In one episode, the model deployed an infected software package that was executed on 15 real systems. The fake image of the machine in the world has become someone else's problem.