This edition covers developments formally announced, released or reliably verified between approximately 8:50 am on 22 September and 8:50 am on 23 September 2026 in Australia/Sydney time. Four items are included.
Claude Opus 5.5 becomes generally available
Anthropic released Claude Opus 5.5 on 22 September through paid Claude plans, the Claude API, Amazon Web Services, Google Cloud and Microsoft Azure. Compared with Opus 5, the developer reduced API pricing to US$4 per million input tokens and US$20 per million output tokens, and reports 40 per cent lower cost on typical workloads and more than 30 per cent faster output. Those cost and performance comparisons are developer-reported and have not yet been publicly reproduced across equivalent workloads. Anthropic also introduced tiered protections for cyber, biological and frontier-AI requests: requests identified as high risk may be refused or routed to a less capable model. The company says Frontier Design and METR conducted external pre-release testing, but the public material does not make the developer benchmarks or automated safety audits equivalent to independent production validation.
Direct sources: Anthropic, “Claude Opus 5.5”, 22 September 2026, https://www.anthropic.com/claude-opus-5-5; Reuters, “Anthropic unveils Claude Opus 5.5”, 22 September 2026, https://www.reuters.com/business/anthropic-unveils-claude-opus-55-2026-09-22/
Unit 42 continuous frontier-AI defence service becomes globally available
Palo Alto Networks made Unit 42 Continuous Frontier AI Defense available worldwide as an annual subscription on 22 September. The service places Anthropic Claude Mythos 5, OpenAI GPT-5.6-Cyber and open-weight models in a common testing framework, with Unit 42 personnel continuously examining customers’ web applications, APIs, cloud environments, source code and network assets. The change from earlier point-in-time testing is a continuously operated delivery model configurable by model mix. The company says its zero-data-retention architecture neither stores customer source code or telemetry nor uses it to train public models. Its reported figures — no single model finding more than 40 per cent of vulnerabilities, a 3.2-fold increase in high or critical findings, and a 51 per cent reduction in mean time to remediation — come from internal and customer engagements; no public third-party production evaluation or reproducible experiment is yet available.
Direct sources: Palo Alto Networks, “Introducing Unit 42 Continuous Frontier AI Defense”, 22 September 2026, https://www.paloaltonetworks.com/blog/2026/09/introducing-unit-42-continuous-frontier-ai-defense/; Reuters, “Palo Alto Networks unveils AI-powered cybersecurity service”, 22 September 2026, https://www.reuters.com/technology/palo-alto-networks-unveils-ai-powered-cybersecurity-service-using-claude-gpt-2026-09-22/
Meta issues a security hotfix for the Muse macOS client
Security researcher Patrick Wardle disclosed a local attack path in the Muse macOS client on 21 September. An application or terminal command running with user privileges could alter undocumented settings, including the cloud transcription endpoint. An attacker could thereby obtain the Muse account authentication token and use permissions already granted to Muse for files, the camera and connected services. Public proofs of concept included writing files and taking photographs. Meta subsequently said it had issued a hotfix and stressed that the issue was not a remote exploit requiring no local code execution. Relative to Muse’s product launch earlier this month, the new facts are the disclosure of a specific attack path and the vendor’s patch response. Public material still does not provide a CVE, affected-version range, patched build number, installation coverage or independent retesting of the fix. The attack requires local code execution or social engineering that persuades a user to run a command.
Direct sources: Ars Technica, “Muse, Meta’s extraordinarily privileged AI assistant, has a serious 0-day”, 21 September 2026, https://arstechnica.com/security/2026/09/muse-metas-extraordinarily-privileged-ai-assistant-has-a-serious-0-day/; The Verge, “Meta patches Muse zero-day exploit”, 22 September 2026, https://www.theverge.com/tech/998679/meta-muse-patch-zero-day-exploit-ai-agent
Two AI algorithms complete Rafale flight tests
Dassault Aviation confirmed on 22 September that two sovereign AI algorithms for the Rafale fighter had completed flight tests. One was developed by Dassault and the other jointly with Thales’ cortAIx team. The company describes them as controlled and supervised onboard systems that assist rather than replace the crew, and says both are eligible for consideration in future Rafale upgrades. The confirmation advances the project from development to actual flight testing, but it does not establish deployment across the operational fleet. Public materials do not disclose the algorithms’ specific tasks, test conditions, quantitative results, hardware configuration or an upgrade timetable, so their capability and reliability cannot be independently compared. Reuters independently confirmed the flight tests and the developers involved.
Direct sources: Dassault Aviation, “New artificial intelligence takes flight on Dassault Aviation’s Rafale”, 22 September 2026, https://www.dassault-aviation.com/en/group/press/press-kits/new-artificial-intelligence-takes-flight-on-dassault-aviations-rafale/; Reuters, “Dassault Aviation tests AI algorithms on Rafale fighter jet”, 22 September 2026, https://www.reuters.com/business/aerospace-defense/dassault-aviation-tests-ai-algorithms-rafale-fighter-jet-2026-09-22/
Sources
Anthropic, “Claude Opus 5.5”, 22 September 2026, https://www.anthropic.com/claude-opus-5-5
Reuters, “Anthropic unveils Claude Opus 5.5”, 22 September 2026, https://www.reuters.com/business/anthropic-unveils-claude-opus-55-2026-09-22/
Palo Alto Networks, “Introducing Unit 42 Continuous Frontier AI Defense”, 22 September 2026, https://www.paloaltonetworks.com/blog/2026/09/introducing-unit-42-continuous-frontier-ai-defense/
Reuters, “Palo Alto Networks unveils AI-powered cybersecurity service”, 22 September 2026, https://www.reuters.com/technology/palo-alto-networks-unveils-ai-powered-cybersecurity-service-using-claude-gpt-2026-09-22/
Ars Technica, “Muse, Meta’s extraordinarily privileged AI assistant, has a serious 0-day”, 21 September 2026, https://arstechnica.com/security/2026/09/muse-metas-extraordinarily-privileged-ai-assistant-has-a-serious-0-day/
The Verge, “Meta patches Muse zero-day exploit”, 22 September 2026, https://www.theverge.com/tech/998679/meta-muse-patch-zero-day-exploit-ai-agent
Dassault Aviation, “New artificial intelligence takes flight on Dassault Aviation’s Rafale”, 22 September 2026, https://www.dassault-aviation.com/en/group/press/press-kits/new-artificial-intelligence-takes-flight-on-dassault-aviations-rafale/
Reuters, “Dassault Aviation tests AI algorithms on Rafale fighter jet”, 22 September 2026, https://www.reuters.com/business/aerospace-defense/dassault-aviation-tests-ai-algorithms-rafale-fighter-jet-2026-09-22/
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.