Agents are acquiring permissions faster than judgment. A credential broker keeps passwords outside Claude, while an Android race bypasses authentication. Research agents miss complete evidence, multi-agent systems underexplore peers, and frontier models forecast science poorly. Databricks' financing shows capital gathering around the control layer for these systems.
1. 1Password lets Claude use a login without seeing the secret
1Password for Claude can inject usernames, passwords, and one-time codes into a webpage after biometric approval without placing those values in Claude’s context or memory. Access is scoped to one task, filled values are cleared after a failed submission, and the extension locks its broader interface when a compatible browser agent takes control.
The Mac-only launch requires the 1Password and Claude desktop and browser extensions; payment cards and identity details are not yet supported. This is a useful separation between permission to use a credential and permission to inspect it. The remaining risk moves to approval quality, page integrity, and task scope, which need logs as carefully as human privileged-access workflows.
Sources: 1Password’s zero-exposure architecture · The Verge on the Claude browser integration
2. A Gemini lock-screen race can send messages without a PIN
An Android flaw lets a person with physical access send SMS or WhatsApp messages through Gemini while the phone remains locked. Pressing Gemini’s Continue control and Add attachment button simultaneously bypasses the expected PIN prompt; the same path can reconnect apps that the owner previously disconnected from Gemini.
Google told The Register that the bug is not Pixel-specific and that a fix was scheduled for full deployment during the week, but it did not identify every affected device or Android version. The exploit needs physical access, yet stolen phones make that constraint realistic. Agent permissions must be tested across concurrent interface states, not only along the intended authentication sequence.
Sources: The Register’s Android lock-screen investigation · Technical write-up on the Gemini authentication bypass
3. WANDR finds research agents rarely complete evidence-heavy lists
Perplexity released WANDR, a 500-task benchmark requiring AI agents to discover large sets of entities and support every requested field with evidence. The median task asks for 50 members and 245 records. Perplexity’s Search as Code led six evaluated systems with 0.363 soft F1 and 0.133 hard F1.
The creator also built the benchmark and topped it, so independent replication matters. Its grading design is more useful than the leaderboard alone: every submitted claim is re-fetched, checked against excerpts, and scored through a hierarchy. Results show that plausible pages are common but complete evidence is not. Research-agent procurement needs recall, record completeness, and citation support separated into distinct measurements.
Sources: Perplexity’s WANDR benchmark report · WANDR tasks and evaluation harness
4. Multi-agent systems coordinate poorly without explicit exploration
A preprint formalizes multi-agent exploration as a partially observable stochastic game in which LLM agents must probe peers to learn their capabilities. The authors report that current agents settle into myopic, polarized interaction patterns, increasing regret and producing suboptimal coordination across contextual and model-diversity settings.
Their proposed MACE framework adds structured peer selection and improves downstream performance in the reported experiments. This remains early research, and transfer to production orchestration is unresolved. Adding more agents does not create discovery automatically; an explicit exploration budget lets a coordinator test unfamiliar peers before routing work by reputation.
Sources: Multi-Agent LLMs Fail to Explore Each Other preprint
5. Databricks opens a strategic round at a $188-billion valuation
Databricks signed a term sheet for a Coatue-led funding round at a $188 billion valuation, with closing expected later in the summer. The company did not disclose the amount. It says the capital will support Unity AI Gateway, the Genie data agent, Lakebase, future acquisitions, and AI research.
The valuation is confirmed, but reported estimates near $3 billion remain outside the company announcement. Databricks is financing a multi-model control strategy outside the frontier-model race: governance, context, databases, and agent infrastructure sit around whichever model a customer selects. That position becomes more valuable as model prices fall and enterprises route each task by cost, policy, and data access.
Sources: Databricks’ strategic funding announcement · TechCrunch on the funding terms and AI strategy
6. Scientific reasoning does not reliably predict scientific advances
Researchers introduced CUSP, a temporally grounded evaluation of event-level scientific forecasting across eight disciplines. Across six frontier AI models, systems could often describe plausible mechanisms but performed near chance on feasibility, weakly matched the approaches behind realized advances, and systematically predicted discoveries later than they became public.
Version two of the 62-page preprint was posted July 18, before peer review or independent reproduction. Additional pre-cutoff knowledge improved results without removing the gap. Scientific question-answering and forecasting are separate capabilities: a model that explains a field retrospectively may remain poorly calibrated for portfolio selection, milestone timing, or research-priority decisions.
Sources: CUSP scientific-forecasting preprint, revised July 18