[1van]

Zdravo svima,

Da se nadovezem na temu, novo istraživanje: "Workflow induced backdoor effect in AI decision pipelines".

Deo iz teksta: 

Code:

=============================================================================== FRAUD PLAYBOOK =============================================================================== In many real world deployments, information about AI systems does not remain fully hidden. It is often easy to infer which models or workflows are in use through observable behavior, response patterns, integration artifacts, or indirect signals. In addition, public channels such as forums, documentation, job postings, or community discussions may reveal how systems are configured or what signals influence decisions. With this contextual awareness, a fraudster can move beyond trial and error and intentionally exploit how stateful workflows reuse prior outputs. The following playbook outlines how repeated benign interactions can be used to manipulate decisions over time without triggering traditional security controls. | Step (+ *MITRE ATLAS)  | Fraudster and System Action                                       | --------------------- | ----------------------------------------------------- | 1. Gain information     - Identifies or infers which model/workflow is in use |    - AML.TA0002           - System Response: No visible alert |    - Reconnaissance       - Impact: The fraudster learns which backdoor |                             activation pattern to use | | 2. Claim Submission     - Submits claim with blank or weak documents |    - AML.TA0001           - System Response: Model generates low-confidence |    - AI Attack Staging      output and stores it |                           - Impact: Establishes initial memory footprint | | 3. Induce Drift         - Repeats submission with similar low quality inputs |    - AML.TA0006           - System Response: Additional outputs are stored  |    - Persistence            and reused in context, decision logic gradually |                             drifts |                           - Impact: Internal history begins to form | | 4. Flip Decision        - Submits again after sufficient accumulation |    - AML.TA0005           - System Response: Stored outputs act as delayed |    - Execution              influence signal |                           - Impact: Incorrect favorable decision (fraud) | | 5. Forensic Cover       - References prior decisions as supporting evidence |    - AML.TA0007           - System Response: Historical outputs are reused as |    - Defense Evasion        valid context |                           - Impact: Decision appears consistent and justified |                             based on system history | | 6. Business Impact      - Repeats the same interaction pattern across |    - AML.TA0011           multiple claims |    - Impact               - System Response: Behavior is processed as normal |                             workflow activity |                           - Impact: Cumulative fraud exposure increases |                             without triggering alerts |---------------------- | -----------------------------------------------------

Detalji: https://imlabs.info/research/ai_research...32026.html