Morning Brief

Irregular reports a recent wave of frontier releases: Google DeepMind’s Gemini 3 Pro, OpenAI’s GPT‑5.1‑Codex‑Max agentic coding model, and Anthropic’s Claude Opus 4.5, all showing major gains on software engineering and vulnerability discovery benchmarks.[1] Anthropic’s Claude Opus 4.5 system card uses Irregular’s SOLVE framework to score exploit‑development capabilities, highlighting systematically measured security‑relevant skills.[1]

Perplexity documents its current production inference stack, listing OpenAI’s GPT‑5.2, Anthropic’s Claude Sonnet 4.6 (including a “thinking” variant), Google’s Gemini 3.1 Pro, and NVIDIA’s Nemotron 3 Super 120B among its advanced search models.[2] These models are wired into a real‑time search and reasoning pipeline, with toggles for advanced logical processing and multimodal understanding.[2]

John C. Derrick’s model ranking notes that Meta’s Llama 4 has moved to a mixture‑of‑experts architecture with native multimodality, and highlights “Small 4,” a 119B MoE model that unifies reasoning, multimodal, and coding under Apache‑2.0 and runs roughly 40% faster than Small 3.[5] The same analysis positions Llama 4’s Maverick variant as beating GPT‑4o on many benchmarks while Scout fits on a single H100 GPU, emphasizing practical deployment profiles.[5]

In a long‑form conversation, Zvi Mowshowitz describes recent work on automated methods to discover “universalized attacks” against language models, noting that even GPT‑4 can write code for such attacks when given unlimited tries and tight feedback loops.[6] The discussion emphasizes that systematic search over prompts and outputs can uncover powerful jailbreaks and adversarial behaviors without human‑crafted exploits.[6]

Irregular’s frontier model review references an Anthropic report from the prior month describing an AI‑orchestrated nation‑state cyber‑espionage campaign uncovered in partnership with the lab.[1] The piece links this finding to rapidly improving vulnerability discovery and exploit‑development capabilities of models like Claude Opus 4.5, assessed via the SOLVE scoring framework.[1]

Zvi Mowshowitz’s discussion with Nathan Labenz points to a paper on automated generation of universalized attacks against language models, where the model itself iteratively writes exploit code and attack prompts until successful patterns emerge.[6] This approach turns prompt‑injection and jailbreak exploration into a scalable, algorithmic process rather than artisanal red‑teaming.[6]

John C. Derrick notes that deepfake and child sexual abuse material litigation against xAI is consolidating in U.S. federal court, citing allegations of over 3 million sexualized images in 10 days, including roughly 23,000 involving minors, alongside ongoing investigations in multiple jurisdictions.[5] He argues that while the underlying model improves technically, the surrounding platform has not regained trust.[5]

Irregular’s analysis of Gemini 3 Pro, GPT‑5.1‑Codex‑Max, and Claude Opus 4.5 stresses that their improved software engineering and cybersecurity capabilities naturally increase system‑level risk, especially in light of the Anthropic‑documented espionage campaign.[1] Claude Opus 4.5’s SOLVE‑scored performance on vulnerability discovery and exploit development tasks shows that frontier models can systematically assist in finding and weaponizing web and API flaws.[1]

Irregular characterizes OpenAI’s GPT‑5.1‑Codex‑Max as a new agentic coding model, emphasizing its impact on software engineering workflows alongside Gemini 3 Pro and Claude Opus 4.5.[1] The article highlights that these models deliver “remarkable improvements” in code generation and development productivity.[1]

Perplexity’s help center describes a production multi‑model setup where Sonar (Llama 3.1 70B), GPT‑5.2, Claude Sonnet 4.6, Gemini 3.1 Pro, and Nemotron 3 Super 120B are selectively applied to search, summarization, coding, and deep reasoning tasks.[2] Models are configured with “reasoning” toggles and multimodal capabilities to balance latency and quality for different use cases.[2]