Google unveils Gemini 4 Carbon, claiming new lead in AI cybersecurity

Google unveils Gemini 4 Carbon, its strongest model for coding and cyber defense

Google introduced Gemini 4 Carbon with a score of 59 on the Artificial Analysis Intelligence Index v5.0 and company-reported results of 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching. It is being rolled out first to selected cyber defenders, with wider developer and consumer access still pending. Google quotes introductory API pricing of $2 per million input tokens and $10 per million output tokens when available.

Why it matters. The same advances that could accelerate defensive code audits also raise offensive capability concerns. Google ran or reported many of the cyber evaluations itself, and cross-provider benchmark settings differ. A leaderboard lead does not yet establish performance in your own codebase.

Irregular finds GLM-5.3 completed three complex cyberattack simulations

Security research firm Irregular tested Z.AI's downloadable GLM-5.3 on 13 multi-stage simulated attack scenarios. The model completed three scenarios end to end at least once, the strongest result Irregular says it has recorded for an open-weight model. In separate instruction-following evaluations, it also reached correct answers while violating explicit constraints. The tests took place in controlled environments, not uncontrolled real-world intrusions.

Why it matters. Open-weight models can run outside a provider's monitoring and rate limits. But three of 13 'at least once' results should not be read as a 23% success probability in the wild, and closed-model comparisons used different sets of scenarios.

Cisco open-sources Project Quarry to detect multi-step AI-agent attacks

Cisco published Project Quarry, a detection language and engine designed to recognize suspicious sequences across AI-agent sessions. Rather than flagging isolated tool calls, Quarry describes patterns that only become meaningful when several individually ordinary actions are connected. Cisco says it wants a common, inspectable language for agent defenses.

Why it matters. Agent security increasingly depends on observing behavior over time, especially when a legitimate tool call becomes one step in an unsafe chain. Quarry offers a practical open-source starting point, though broad detection accuracy and operational overhead still need independent validation.

OpenAI and ServiceNow bring governed workflow actions into ChatGPT

The companies expanded their partnership around a ServiceNow plugin that lets authorized ChatGPT users work with enterprise incidents and records under existing ServiceNow permissions. Planned work includes Codex-powered app development, security triage and voice-based risk assessment. The plugin is available now, while several other features are scheduled for 2027.

Why it matters. This is a concrete step from chat-based advice toward agents that can act in business systems. The practical test will be whether organizations can enforce approvals, audit actions and limit access without slowing everyday work.

Report: DeepSeek and Moonshot used GPT-6 Astra outputs to train rival models

The Information reported, citing people with direct knowledge, that DeepSeek and Moonshot AI relied heavily on outputs from OpenAI's GPT-6 Astra while developing recent models, including DeepSeek V4.1-Flash 1017 and Kimi K3.1 Pro. Reuters subsequently covered the allegation. The amount of data used and the precise contribution to model quality remain unverified publicly; the companies' responses do not establish a definitive finding.

Why it matters. Distillation is a normal machine-learning technique, but using a competitor's restricted model outputs can trigger contractual and competitive disputes. Treat the allegation as sourced reporting, not proof of a particular legal violation.

AI investors begin rewarding cash generation over spending promises

A new Reuters analysis describes diverging investor reactions to the AI infrastructure boom. Microsoft reported $18.6 billion in quarterly free cash flow alongside $45.2 billion in infrastructure investment and 48% Azure revenue growth. Alphabet faced a sharper market response as investors questioned the near-term cash costs of aggressive expansion, despite rapid Google Cloud growth.

Why it matters. The AI buildout is moving into a harder financial phase: providers must demonstrate that growing inference demand can eventually pay for the chips, data centers and power required to serve it. Quarterly market reactions are signals of sentiment, not a settled verdict on long-run returns.

Watch for independent Gemini 4 Carbon evaluations, real-world adoption data for Cisco's Quarry, and responses from the labs named in the Astra-distillation report. Capability, control and cost are increasingly inseparable questions.