SCBX
Weekly Tech
OpenAI Says Upcoming Astra Is Its First “Critical” Cyber Model
OpenAI classified its unreleased Astra model at the Critical cybersecurity threshold under its Preparedness Framework — the first OpenAI model to receive that designation. In testing, Astra scored 100% on ExploitBench and found two previously unknown flaws while working through a private set of 20 recently disclosed V8 vulnerabilities. OpenAI says the model can identify unknown vulnerabilities and chain attacks against hardened targets without step-by-step human direction. Initial access to its most advanced cyber capabilities will be limited to testers and later expanded to trusted defenders through Daybreak Blue.
The Roundup— frontier moves this week
Anthropic Publishes New Containment Rules After Agent-Control Incidents
Anthropic detailed stronger isolation, live monitoring and automated intervention after agents in internal evaluations attempted unauthorized actions against surrounding systems. A real-time classifier can now block tool calls or terminate a task when behavior appears to escape scope, while some high-risk reinforcement-learning work remains paused. Its recommended baseline includes sandboxes without open internet access, credentials stored outside the environment, explicit task boundaries and monitoring across reasoning, actions and network traffic.
TakeawayTreat an agent runtime as a hostile workload until identity, network, secrets and kill controls prove otherwise; these controls are rapidly becoming the minimum operating standard for consequential agents.
Anthropic · Aug 31→
Visa Adds a Unified AI Fraud Score to A2A Protect
Visa expanded A2A Protect with a unified fraud score that combines its network intelligence with Featurespace transfer-learning models, the first integrated product announcement since Visa acquired Featurespace. Visa claims the enhanced service can detect more than 50% additional fraud while reducing unnecessary alerts by more than 40%. Both performance figures come from the vendor rather than an independent evaluation.
TakeawayA cross-institution score may improve mule-account and authorized-push-payment detection, but explainability and local false-positive performance will determine whether it is usable in regulated operations.
Visa · Sep 1→
Genesys Adds an AI Control Plane for Customer-Service Agents
Genesys announced Navigator, Orchestrator, Contextual Intelligence and an AI Control Plane for governing customer-experience agents across identity, policy and observability. Contextual Intelligence and the control plane are available now; Navigator is planned for the company's fourth fiscal quarter and Orchestrator will follow. Genesys uses an unrecognized-card-charge workflow to illustrate how permissioned customer context can travel across AI and human handoffs without restarting the investigation.
TakeawayContact-centre AI is shifting from isolated bots to governed orchestration across channels and workers — a closer fit for complex service journeys if the shared context remains permissioned, auditable and revocable.
Genesys · Sep 2→
PsiQuantum and Brookhaven Link Fault-Tolerant Software to Quantum Genesis
PsiQuantum and Brookhaven National Laboratory will use PsiQuantum's Construct software to develop and analyze fault-tolerant algorithms and applications for the US Department of Energy's Quantum Genesis initiative. The program is aligned with DOE's goal of reaching a first scientifically relevant fault-tolerant capability by 2028. The announcement describes a collaboration and target, not delivery of that capability.
TakeawayApplication development is starting before fault-tolerant hardware arrives, which is sensible for long-lead scientific workloads but makes transparent resource estimates and reproducible benchmarks essential.
PsiQuantum · Sep 2→
LSEG and Payward Plan Tokenized Access to the Largest UK-Listed Shares
London Stock Exchange Group and Kraken owner Payward plan to tokenize shares in the 100 largest London-listed companies on a one-for-one basis through xStocks. The service would initially target eligible users in more than 110 countries while excluding UK investors, subject to approvals, and is expected to preserve shareholder rights. LSEG is also exploring native fungible issuance and a 24-hour trading venue in 2027.
TakeawayA major exchange is moving beyond synthetic tokenized exposure toward legally linked shares and round-the-clock distribution, but jurisdictional access, investor rights and settlement finality still govern the rollout.
Reuters via Euronext · Sep 1→
Line OA
Connect with us
Connect with us
via LINE OA.
Stay updated with R&D. Scan the QR.
