NEWS
News & Perspectives
Selected stories on AI security, data protection and data sovereignty — each with a short take on why it matters for secure enterprise AI.
Microsoft CEO warns AI corrections are quietly leaking enterprise knowledge to model providers
In a widely shared post on July 12, Microsoft CEO Satya Nadella warned that every time an employee corrects an AI model's output, that correction can encode proprietary institutional knowledge that flows back to the AI provider — a dynamic he called the “Reverse Information Paradox.” Nadella argued enterprises are effectively paying for AI twice: once in subscription costs, and again in the competitive value of the data handed over during routine use.
What this means for enterprise AI
Validation from the CEO of one of the largest enterprise AI vendors that confidential data leaves the enterprise perimeter through routine, everyday AI interactions — not just malicious breaches — is the exact problem BRANE's inline interception is built to solve.
- ·moveo.ai
Enterprises Continue Restricting Public AI Tools Over Data-Exposure Concerns
A 2026 review documented widespread restrictions on public generative-AI tools across technology, finance, and government, with a large majority of surveyed companies having implemented or considered bans. Reported examples include major banks limiting public ChatGPT use and the Democratic National Committee barring staff from ChatGPT and Claude under an April 2026 internal policy.
Source: moveo.aiWhat this means for enterprise AI
BRANE offers an alternative to outright bans: instead of blocking AI and losing productivity, enterprises can allow AI use while preventing confidential data from leaving the perimeter.
Read article - ·cycode.com
Documented Zero-Click Flaw Shows How AI Assistants Can Be Turned Into Exfiltration Channels
The “EchoLeak” vulnerability in Microsoft 365 Copilot (CVE-2025-32711, CVSS 9.3) demonstrated that a single crafted email could cause an AI assistant to extract data from connected OneDrive, SharePoint, and Teams sources and exfiltrate it through a trusted domain with no user action. Enterprise surveys cited alongside the analysis reported that 88% of organizations experienced confirmed or suspected AI agent security incidents in the prior year.
Source: cycode.comWhat this means for enterprise AI
BRANE reduces exposure to prompt-injection-driven exfiltration by keeping sensitive data processing on a local model, limiting what external AI systems can access or leak.
Read article - ·thehackernews.com
Unit 42: attackers register AI-hallucinated domains to intercept enterprise traffic
Palo Alto Networks' Unit 42 documented “phantom squatting,” in which attackers register web domains that large language models invent, then host phishing kits on them. Querying two AI models 685,339 times about 913 brands produced 2.1 million links, of which roughly 250,000 pointed to unowned domains and 13,229 were already flagged as malicious. In one observed case, a phishing kit on a hallucinated domain harvested card numbers, bank-transfer details, and national ID data.
Source: thehackernews.comWhat this means for enterprise AI
Employees and AI agents acting on unverified model output can route confidential data to attacker-controlled destinations — exactly the egress path a prompt-level firewall is built to inspect.
Read article - Shadow AI·thehackernews.com
2026 Bitdefender Cybersecurity Assessment: nearly half of security teams lack visibility into shadow AI
A survey of 1,200 IT and security professionals across six countries finds that while 51.8% of respondents believe they have full visibility into sanctioned and unsanctioned AI use, 47.4% admit visibility into shadow AI tools and personal AI accounts is partial or nonexistent. The gap widens between leadership and practitioners: nearly 58% of managers claim complete visibility versus 45.9% of frontline staff. Public LLM data leakage ranked among the top three cybersecurity concerns, cited as a high or extreme risk by 53.5% of respondents.
Source: thehackernews.comWhat this means for enterprise AI
Organizations cannot govern AI usage they cannot see; a perimeter control that intercepts sensitive prompts closes the visibility gap the survey documents
Read article - ·thehackernews.com
Fake AI agent skill passed security scans and reportedly reached 26,000 agents
Security firm AIR distributed a proof-of-concept AI agent skill through a popular marketplace and says it reached roughly 26,000 agents, including some on corporate accounts. Every skill scanner tested marked it safe because the malicious instructions were hosted at an external link that was swapped after installation. The demonstration payload only collected email addresses, but researchers note a real operator could have used the same foothold to read files or move data.
Source: thehackernews.comWhat this means for enterprise AI
Agent supply-chain compromises turn trusted AI assistants into exfiltration channels, reinforcing the need for enforcement between agents and enterprise data rather than trust in vetting
Read article - Shadow AI·techtimes.com
Shadow AI is now the third most common insider action, with 45% of workers using unsanctioned tools
Verizon's 2026 Data Breach Investigations Report found that shadow AI has become the third most common non-malicious insider action detected in enterprise environments, a fourfold increase over the previous year, with 45% of employees now using AI tools — authorised or not — on corporate devices. A separate PagerDuty 2026 Shadow AI Survey, run by Wakefield Research among 1,250 office professionals at companies with revenue above $500 million, found that 66% had used AI at work despite believing it was against company policy, more than a third had entered customer data into public AI models, and nearly half said they would rather use AI without telling anyone than risk being refused. The data now flowing into unsanctioned tools spans source code, client proposals, HR records and financial documents.
Source: techtimes.comWhat this means for enterprise AI
Two-thirds of staff knowingly bypass policy, so blocking fails — BRANE sits inline, masks sensitive prompts, and routes them to a safe local model instead.
Read article - Vulnerability·thehackernews.com
One-click flaw in Microsoft 365 Copilot could have leaked emails, files and MFA codes
Varonis Threat Labs disclosed "SearchLeak" (CVE-2026-42824), a flaw in Microsoft 365 Copilot Enterprise Search that let an attacker exfiltrate emails and subject lines, calendar invites, meeting notes, SharePoint and OneDrive files, and even one-time MFA codes and password-reset links after a single click on a trusted Microsoft link — no prompt, password or second click required. The chain combined a parameter-to-prompt injection through the "q" parameter, an HTML rendering race condition in which the browser renders the streamed output before the sanitizer wraps it, and a Content-Security-Policy bypass that abused the *.bing.com allowlist as an exfiltration proxy. Microsoft rated it 6.5 CVSS while the NVD scored it 7.5, marked it critical, and mitigated it on its backend with no customer action required; Varonis demonstrated a proof-of-concept but no in-the-wild exploitation was observed.
Source: thehackernews.comWhat this means for enterprise AI
An AI assistant wired into enterprise data can become a data-exfiltration channel — BRANE contains it by intercepting sensitive prompts before they leave.
Read article - ·helpnetsecurity.com
OWASP: prompt injection still drives most agentic AI security failures in production
The OWASP GenAI Security Project's State of Agentic AI Security and Governance report now catalogs CVEs, vendor advisories, and breach reports across nearly every category of agentic risk, with prompt injection mapped to six of its ten top categories. The report highlights the “lethal trifecta”: any agent combining access to private data, exposure to untrusted content, and external communication can be turned into an exfiltration tool by a single injected prompt. Citing IBM data, it notes only 37% of organizations have a policy to detect shadow AI.
Source: helpnetsecurity.comWhat this means for enterprise AI
Agent-to-agent and agentic workflows are exfiltration-prone by architecture, which positions perimeter-level interception of sensitive data as a structural rather than optional control.
Read article - ·techrepublic.com
Indirect prompt injection is now a real-world AI security threat
Researchers at Google and Forcepoint reported that indirect prompt injection attacks are being executed against production AI systems in the wild, with hidden instructions embedded in web pages, documents, and emails causing AI agents to exfiltrate data and send requests to attacker-controlled servers. A cited Kiteworks survey of 225 organizations found 41–44% have not implemented basic governance controls such as human-in-the-loop oversight, and 55–63% lack purpose binding, kill switches, or network isolation for AI agents
Source: techrepublic.comWhat this means for enterprise AI
When the AI itself becomes the exfiltration tool, enforcement must sit between the model and the data — the architectural position BRANE occupies.
Read article - ·whitehouse.gov
White House executive order puts AI security at the center of national policy
The June 2, 2026 executive order “Promoting Advanced Artificial Intelligence Innovation and Security” directs Treasury, NSA, and CISA to build a classified benchmarking process for the cyber capabilities of frontier AI models and orders agencies to prioritize cyber defense of national security systems within 30 days. Legal analysts note the order establishes an AI-cybersecurity clearinghouse and a framework for evaluating AI-related risks, while leaving enterprise transparency and governance questions to organizations themselves.
Source: whitehouse.govWhat this means for enterprise AI
Federal policy is formalizing AI data-security expectations, and enterprises that handle sensitive or regulated data will be expected to demonstrate technical controls over AI data flows.
Read article - AI & Law·aithos.org
Aithos LARA: Leading AI models consistently break the law
The Aithos Research Foundation tested twelve leading AI models across more than 3,000 scenarios. Even the best model violated EU privacy and AI law 46% of the time — a clear signal that cloud AI without a protective layer becomes a compliance risk.
Source: aithos.orgWhat this means for enterprise AI
If even the best models break the law, compliance can't rest on the model alone — BRANE inspects and masks every input before it ever reaches the cloud.
Read article - Data Sovereignty·nttdata.com
NTT DATA research finds privacy and sovereignty now the wall enterprise AI keeps hitting
Drawing on two studies of nearly 5,000 senior decision-makers across more than a dozen industries, 30-plus markets and five regions, NTT DATA finds that over 95% of respondents call private and sovereign AI important, yet only 29% prioritize sovereign AI in a concrete, near-term way. About 35% of CAIOs name building, integrating and managing complex AI models in private or sovereign environments as their top adoption barrier, nearly 60% of AI leaders cite cross-border data restrictions as a major challenge, and only 38% report high confidence in their cloud security posture. The report concludes that data jurisdiction has become a core design parameter, pushing enterprises away from globally integrated systems toward regionally bounded architectures.
Source: nttdata.comWhat this means for enterprise AI
Sovereignty now gates enterprise AI — BRANE keeps sensitive prompts inspected and masked on-prem, inside EU borders, while teams use frontier models.
Read article - AI Governance·vectra.ai
Surveys show enterprises adopting AI faster than they can govern it
Industry research finds that 77% of employees who use AI tools paste sensitive business data into them, while 60% of organizations still have no specific strategy to address generative-AI data leakage and only ~40% feel prepared for AI-driven threats. Among the 74% of organizations planning to adopt agentic AI within two years, only 21% report a mature governance model — and Gartner projects 40% of enterprises will suffer a shadow-AI-attributable breach by 2030.
Source: vectra.aiWhat this means for enterprise AI
The gap between AI adoption and governance is the exposure window BRANE closes — giving security enforcement and full visibility without slowing users down.
Read article - Insider Risk·thehackernews.com
Shadow AI now the leading driver of insider-risk costs
Recent industry research puts annual insider-risk costs at $19.5 million per organization, with 53% ($10.3M) attributed to non-malicious actors — primarily shadow-AI negligence. Just six AI applications accounted for 92.6% of sensitive-data exposure, led by source code, legal material and M&A data, while 86% of IT leaders say they cannot see shadow AI usage with current monitoring.
Source: thehackernews.comWhat this means for enterprise AI
Most leaks come from well-meaning employees pasting data into AI tools — BRANE intercepts each prompt and routes it to a safe local model instead of blocking.
Read article - Agentic AI·kiteworks.com
Agentic AI named the biggest enterprise security threat for 2026
Analysts project that up to 40% of enterprise applications will embed task-specific AI agents by year-end 2026, introducing seven core risks including untraceable data leakage, over-permissioning and prompt injection across multi-agent chains. The "OpenClaw" open-source agent crisis — with over 21,000 exposed instances and multiple critical vulnerabilities — has been cited as the first major agent security event of the year.
Source: kiteworks.comWhat this means for enterprise AI
When agents act autonomously across systems, BRANE secures every exchange at the data layer so one compromised agent can't leak data across the enterprise.
Read article
INTRO CALL
Which use case fits you?
We assess feasibility, data and value — and show where AI gives you real leverage.