Attack
According to Wiz, the public reports describe a targeted search for environment variables giving access to the AI.
One key was used for about 1.7 million calls to Bedrock in 13 hours, of which more than a million succeeded.
Actors linked to LLM resale sites tested both keys, with no established link to their theft.
News
More than 200 organizations in at least 14 countries were targeted; nine experienced post-compromise activity.
ARTEX coordinated agents with DeepSeek as the main model; human operators directed their work.
An unsigned SAML assertion and an XML external entity flaw were used to obtain cloud credentials.
Oktsec recommends verifying the rejection of unsigned SAML assertions and the absence of a session after rejection.
News
Agents can act with a person's credentials, with their actions attributed to that person.
The principles plan to distinguish each agent from its operator and to limit its permissions to its task.
According to the author, access control must stay outside the model's reasoning, which can be influenced by untrusted content.
According to NIST, current standards and technologies do not yet cover all the principles in all deployments.