CodeWall finds flaw in McKinsey’s Lilli as Google Maps adds Ask Maps

Security company CodeWall published an account on March 9 of its test of McKinsey’s internal AI assistant, Lilli. It said its autonomous agent gained read-and-write access to Lilli’s production database in about two hours. CodeWall described a path through an unauthenticated interface and SQL injection. It also said the agent could reach system instructions and configuration data.¹²

McKinsey confirmed a vulnerability and said it fixed the issue within hours. An investigation supported by outside forensic specialists found no evidence that the researcher or any unauthorized third party accessed client data or confidential client information, the firm said.³ The case illustrates a risk for internal AI systems: weak interfaces and database permissions can expose more than individual chat responses. CodeWall’s account of what its agent could access should not be mistaken for evidence that client data was taken.¹³

OpenAI announced new capabilities for its Sora 2-powered Video API on March 12, 2026. Developers can use custom characters and objects as references, continue existing videos, and submit multiple video generations as batch jobs. OpenAI says the features are available to all Video API customers. Video continuation is intended to let a scene extend beyond a single generated clip.¹²

The update supports workflows that build videos in several connected steps rather than treating each clip as a separate task. Character and object references give developers a way to carry visual inputs across generations.¹² OpenAI did not specify how developers can persistently manage those references. It also did not quantify time savings or throughput gains for batch jobs.¹²

On March 12, 2026, Google announced two new features for Google Maps. Ask Maps uses Gemini to answer questions about places in a conversation, including requests to find stops for a trip.¹² It is rolling out first on Android and iOS in the U.S. and India. Immersive Navigation is launching first in the U.S. and presents driving routes in a 3D view.¹

The navigation view shows buildings, overpasses, and terrain while highlighting road details.¹ The update pairs a more conversational way to find places with additional visual cues for decisions while driving.¹ For businesses with physical locations, Ask Maps could affect how people discover places. Google has not given a firm timetable for availability beyond the announced markets.¹

Anthropic announced the Claude Partner Network on March 12, 2026. The free program is open to organizations that bring Claude to enterprises, and applications opened with the announcement.¹ Anthropic has committed an initial $100 million for 2026 to support training, technical assistance, and joint market development. Partners also gain access to sales and marketing resources.¹

The network expands Anthropic’s work with outside service providers that help companies adopt Claude. Accenture, Deloitte, Cognizant, and Infosys have been named among its early partners.²³ The program is not limited to consulting firms: Anthropic says any organization bringing Claude to market can apply.¹ It has not specified how much of the commitment will go directly to partners rather than other forms of support.¹²

Anthropic announced a beta feature for Claude on March 12, 2026: The assistant can create interactive charts, diagrams, and visualizations directly in a chat.¹ They appear within the conversation rather than only in a separate view such as Claude Artifacts.¹ Anthropic says the feature is available on every plan, including the free tier.¹

Unlike Artifacts, the new visuals are not intended as persistent workspaces: They can change as the conversation continues.¹² That lets users explore questions about data or processes visually without switching views.¹ These are interactive elements, not generated images, and Anthropic is introducing the capability as a beta feature.¹³

Cursor introduced CursorBench on March 11, 2026. The internal evaluation compares AI coding agents using tasks taken from real engineering sessions through Cursor Blame and prepared for testing.¹ Prompts are deliberately short and underspecified to resemble everyday developer requests rather than carefully written test questions.¹ The test set is not publicly available.¹²

Cursor says public coding benchmarks can struggle to reflect its users’ work, handle ambiguous grading, and avoid potential overlap with training data.¹ It therefore checks results from controlled tests against evaluations from actual usage.¹ For companies, the approach highlights why a strong score on standardized coding tasks does not necessarily mean an agent will make reliable changes to their own code. Because the tasks and grading are not open, CursorBench results cannot be independently verified.¹²