Standardized benchmarks that rank LLM performance by rewarding confident correct answers and penalizing expressions of uncertainty create incentives for models to produce confident guesses rather than explicitly indicating uncertainty or saying 'I don't know'.
high
mechanism
Benchmark-driven evaluation criteria influence model behavior by shaping the objectives optimized during fine-tuning and deployment.
Malware variants have been observed using large language models (LLMs) to change behavior mid-attack, enabling dynamic generation of malicious scripts, on-demand creation of malicious functions, and code obfuscation to evade detection.
high
technical
Describes capabilities attributed to emerging AI-enabled malware that leverage LLMs during active intrusions.
Some malware can call out to LLMs (including proprietary models such as Gemini) to rewrite their own source code, disguise malicious activity, and attempt lateral movement across connected systems.
high
technical
Refers to use of external AI models to modify malware behavior and aid persistence and propagation within networks.
Blackwell is Nvidia's proprietary graphics processing unit (GPU) architecture that is used to power large language models.
high
definition
Describes the Blackwell GPU family and its role in training and inference for large language models.