AI security blind spots, unmonitored code risks, and cost-control strategies

Tools and stories covered

1. 1. VentureBeat

   Prompt injection ranks No. 1 with OWASP and No. 12 in the incident record. The attack itself is invisible to a scan.

   Summary

   A review of over 6,600 real-world security incidents showed that prompt injection attacks ranked twelfth in reported occurrences, despite being listed as the top security vulnerability for large language models by OWASP for three consecutive years. The discrepancy occurs because standard automated vulnerability scanners cannot detect these attacks, which manipulate model instructions rather than exploiting traditional code flaws. Enterprise chief information security officers and IT risk managers must adjust auditing strategies beyond standard software scanning.

   Why this matters

   Mid-sized businesses relying on standard vulnerability scans may be underestimating their exposure to AI manipulation and need dedicated testing processes for conversational tools.

   Read the original on VentureBeat

2. 2. CIO Dive

   Why wild code is an IT crisis hiding in plain sight

   Summary

   The widespread availability of generative AI tools has enabled employees outside IT departments to generate functional software and automation scripts without central approval or review. This unmonitored code introduces governance, operational, and data security risks as business operations begin depending on informal applications. IT managers and department leads need to establish clear software oversight standards that account for AI-assisted development across all business units.

   Why this matters

   Operations leads should establish clear policies and inventory processes for employee-created AI scripts before informal tools become critical to business workflows.

   Read the original on CIO Dive

3. 3. CIO Dive

   Travelers builds its own LLM, cutting AI costs

   Summary

   Insurance firm Travelers developed its own specialized internal model to process industry-specific insurance queries directly. The organization reserves larger, general-purpose third-party models exclusively for broader reasoning, research, and technical coding tasks. This hybrid approach allows company leaders to lower operating expenses by steering routine domain tasks to smaller, dedicated tools.

   Why this matters

   For mid-sized companies facing high AI usage costs, routing routine business tasks to smaller or specialized models can help keep operating budgets under control.

   Read the original on CIO Dive

4. 4. TechCrunch AI

   Claude Cowork finally remembers what you told the app in chat

   Summary

   Anthropic has updated Claude to share memory between its standard chat interface and its Cowork workspace environment. The change allows the assistant to retain project background, user preferences, and business context across different tasks without requiring users to repeat instructions. Department managers and team members using the tool can work across projects with less setup overhead.

   Why this matters

   This update reduces repetitive administrative effort for teams using Claude across multiple administrative and coordination workflows.

   Read the original on TechCrunch AI

5. 5. VentureBeat

   Perplexity partners with Nvidia to launch Portable Computer, a fully local AI agent with zero token costs

   Summary

   Perplexity and Nvidia have launched Portable Computer, an software setup designed to run AI agent tasks locally on owned hardware, starting with desktop supercomputers and Linux machines equipped with Nvidia RTX graphics processors. By processing models and files entirely on local devices rather than in the cloud, the system eliminates recurring cloud token processing costs. IT directors evaluating hardware options can consider this model for handling local data processing requirements.

   Why this matters

   Companies with strict cloud cost constraints or specialized local hardware can consider local model execution to eliminate recurring pay-per-use cloud processing fees.

   Read the original on VentureBeat