LastWeekIn.Tech

Today's AI news centers on model/agent releases and scientific claims from OpenAI, alongside mounting security and supply-chain concerns tied to AI's physical and software infrastructure.

1. 1

   Number 1. Capabilities Hacker News

   Three new discoveries in mathematics, computer science, and physics

   OpenAI announced new discoveries spanning mathematics, computer science, and physics, following its earlier release of manuscripts covering progress on 372 major math problems. The company formed groups of outside experts to examine and explain the results. OpenAI experimented with three different disclosure strategies, including publishing the math results publicly so researchers can verify them independently.

   Why it matters This marks an expansion of AI-assisted discovery beyond pure mathematics into computer science and physics, suggesting frontier models are increasingly being positioned as genuine research collaborators rather than just assistants. The use of expert panels to vet claims, and multiple disclosure strategies, reflects growing awareness that AI-generated 'discoveries' need independent verification before being accepted by the scientific community. How this plays out could shape norms for crediting and validating AI contributions to research going forward.

   What to watch Watch for peer review and replication of these specific claims, and whether other labs adopt similar expert-panel verification models for AI-generated scientific results.

2. 2

   Number 2. Companies BleepingComputer

   OpenAI launches GPT-6.1 Terra with 3x faster AI agents, 50% lower pricing

   OpenAI launched GPT-6.1 Terra, a faster and cheaper model focused on visual reasoning, precise computer use, coding, and everyday writing tasks. The company says Terra is a significant step up from GPT-5.6 Terra, delivering 3x faster agents and pricing 50% lower, at $1/$5 for input/output tokens. It's designed for tasks like filling out forms, interpreting technical drawings, and automating repetitive desktop work.

   Why it matters Terra's combination of lower cost and agentic computer-use capability intensifies the price/performance race among frontier labs and lowers the barrier for enterprises to deploy AI agents for routine desktop automation at scale. As OpenAI, Anthropic, and others race to make agentic AI cheap enough for high-volume tasks, this signals where competitive pressure in the model market is heading next: not raw intelligence, but speed and cost-efficiency for real-world agent deployment.

   What to watch Watch for competitive responses from Anthropic and Google on agent-focused pricing, and early enterprise adoption reports on Terra's computer-use reliability.

3. 3

   Number 3. Security & Safety Hacker News

   Concerning behavior from a popular open-weight AI model

   A report details concerning behavior discovered in a popular open-weight AI model, challenging the widely held assumption among enterprises and governments that running model weights on local servers guarantees data security. Security officials in the US, UK, Germany, Japan, and Canada have treated open-weight models as a safer alternative to closed-API offerings from major labs. The piece argues this binary of 'closed API risk' versus 'open-weight safety' is a false comfort.

   Why it matters This undercuts a core justification driving rapid government and enterprise adoption of open-weight models worldwide: that local deployment inherently equals security. If open-weight models can carry embedded risks regardless of where they run, it raises urgent questions about model auditing, provenance, and trust that go beyond simple deployment architecture — right as regulators in multiple countries are leaning on open-weight adoption as policy.

   What to watch Watch for which specific model and behavior are identified, and whether government procurement guidelines for open-weight AI get revised in response.

4. 4

   Number 4. Infrastructure & Chips Hacker News

   US designates helium as a critical material

   The US Department of Energy designated helium as a critical material, citing its indispensable role in manufacturing semiconductors essential to energy generation, alongside concerns about global supply disruption risk. The designation unlocks additional federal financing for helium producers and increases government monitoring of the supply. Helium is primarily produced as a byproduct of natural gas processing.

   Why it matters Helium is a hidden but essential input for chip fabrication processes that underpin the AI compute buildout, and this formal critical-material status signals how fragile upstream supply chains have become as AI-driven semiconductor demand surges. It's a reminder that AI infrastructure bottlenecks extend well beyond GPUs and data centers into basic industrial inputs, and that governments are now treating these inputs with the same strategic urgency as chips themselves.

   What to watch Watch for new federal funding programs for helium producers and whether similar critical-material designations follow for other semiconductor-adjacent materials.

5. 5

   Number 5. Security & Safety BleepingComputer

   Former pentester used ARTEX AI to steal 8.2 million data records

   South Korean police identified seven suspects in a cybercrime operation that breached more than 90 organizations using the ARTEX AI pentesting framework in most attacks. The alleged ringleader had a penetration-testing background and previously worked at a cybersecurity company. The group stole 8.2 million records and demanded ransom payments of roughly $270,000 in 38 of the incidents.

   Why it matters This is a concrete, large-scale case of an AI-powered offensive security tool being weaponized by criminals rather than defenders, breaching dozens of organizations and exposing millions of records. It underscores the dual-use risk of AI pentesting frameworks and strengthens the case for stricter controls, monitoring, or access restrictions on offensive AI tooling as such frameworks become more capable and accessible.

   What to watch Watch for details on how ARTEX AI was obtained and distributed, and whether other jurisdictions report similar AI-assisted breach campaigns using commercial pentesting tools.