# The Week in Cyber: 7.6 Billion Phishing Emails, New Passkey Attacks, and the Rise of Agentic AI Threats

In a week packed with critical cybersecurity developments, researchers uncovered alarming trends in AI-powered social engineering, demonstrated new vulnerabilities in passkey authentication, and revealed that phishing remains the dominant entry point for devastating data breaches. With Microsoft detecting a staggering 7.6 billion phishing emails in Q2 2026 alone, the threat landscape has never been more hostile. Let's dive into the most significant findings that every security researcher and ethical hacker needs to know.

## OWASP's 2026 LLM Top 10: Prompt Injection Still Reigns Supreme

The Open Worldwide Application Security Project (OWASP) has released its updated Top 10 security risks for Large Language Model applications, and the results confirm what many vulnerability researchers have long suspected. Prompt injection retains its position as the number one threat, but the 2026 update introduces a new concern: hidden context exposure.

This emerging vulnerability class exploits the way LLMs process information embedded in system prompts or training data that users never see. For penetration testers and security teams, this means the attack surface for AI applications continues to expand beyond traditional input validation. Organizations deploying LLM-powered tools must now account for the possibility that sensitive information buried in model context could leak through carefully crafted queries.

The OWASP update underscores a broader truth about modern cybersecurity: traditional vulnerability assessment approaches are insufficient when dealing with AI systems. Security professionals need to develop new methodologies for testing LLM implementations, and the fact that hidden context exposure is now recognized as a distinct risk category suggests we're only scratching the surface of AI-specific attack vectors.

## Passkey Authentication Under Fire: Three New Attack Families Revealed

Black Hat researchers presented findings that challenge the assumed security of passkey authentication systems. The new research unveiled three distinct passkey attack families that manage to bypass authentication logging and synchronization mechanisms. This is particularly concerning because passkeys have been widely promoted as the future of passwordless authentication, a supposedly more secure alternative to traditional credentials that are so often compromised in data breaches.

The vulnerabilities discovered by researchers demonstrate that passkey implementations carry their own set of security risks. Authentication logging bypass means that attackers can potentially gain unauthorized access without leaving forensic traces, making incident response significantly more difficult. The sync-related attacks are especially worrying for enterprises that have embraced passkey-based single sign-on solutions across their infrastructure.

For security researchers, these findings serve as a reminder that no authentication method is foolproof. The race between security controls and the attackers who seek to subvert them continues, and passkey systems now find themselves under the microscope. Organizations that have already deployed passkey technology should monitor vendor patches closely and reevaluate their authentication security posture.

## The Phishing Epidemic: Numbers That Demand Attention

The statistics coming out of Q2 2026 are nothing short of alarming. Microsoft alone observed approximately 7.6 billion phishing emails during the quarter, a number that seems almost incomprehensible in its scale. When you factor in that phishing remains the primary initial access vector in over half of all attacks observed by Cisco Talos, the picture becomes clear: email security is the frontline of organizational defense.

KnowBe4's coverage of these findings highlights an uncomfortable truth - human error remains at the core of AI-enabled social engineering. The Resilience data shows that AI-enabled social engineering drove a staggering 85% of cyber insurance losses in the first half of 2026. This isn't just a technical problem; it's a business continuity crisis.

What makes these phishing campaigns so effective is the growing sophistication of AI-generated content. KnowBe4 reports that AI chatbots are actually outperforming human scammers at building trust in pig-butchering scams, a form of long-game social engineering that involves building relationships with victims over extended periods. The implications for security awareness training are profound - if AI can be more persuasive than trained social engineers, traditional cybersecurity education approaches may need significant revision.

## Zero-Day Vulnerabilities: Akamai's Pre-Disclosure Defense Strategy

In the high-stakes world of zero-day exploits, Akamai Technologies is taking an innovative approach to defending its clients before the broader industry even knows about the vulnerability. Their strategy involves staged rollouts and shared intelligence across their global platform, allowing them to identify and mitigate potential threats based on behavioral patterns rather than known signatures.

This proactive stance on vulnerability management acknowledges a harsh reality of the modern threat landscape: time-to-exploitation has gone negative, meaning attackers are weaponizing vulnerabilities before patches are even developed. Cisco has echoed this concern, arguing that the shrinking patch window demands a new operational approach they call "VulnOps," with managed service providers running network operations centers to close the gap.

For organizations without access to the intelligence-sharing networks that companies like Akamai and Cisco maintain, the situation is even more dire. The traditional model of waiting for responsible disclosure followed by patching is no longer viable when sophisticated threat actors are already deploying exploits during what should be the safe window.

## JavaScript Obfuscation: The Evolution of Web-Based Attacks

Cisco Talos has published research examining JavaScript obfuscation and its transformation from a developer optimization technique into a staple of phishing kits and malware distribution campaigns. What began as a method to protect intellectual property or reduce file sizes has become a cat-and-mouse game between attackers seeking to evade detection and security researchers working to reverse these increasingly complex obfuscation techniques.

The research highlights several approaches to defeating JavaScript obfuscation, including static analysis, dynamic instrumentation, and the use of headless browsers to observe behavior in safe environments. For security teams monitoring web traffic, understanding these obfuscation techniques is essential for identifying malicious scripts hiding in otherwise legitimate-looking web pages.

## Agentic AI Security: The New Frontier

As enterprises race to adopt agentic AI systems, a significant gap has emerged between adoption speed and security readiness. Deloitte, KPMG, and Accenture surveys all arrive at the same uncomfortable conclusion: organizations are deploying AI agents faster than they're building the governance structures needed to secure them.

Cisco is arguing that the answer lies in the Agentic SOC (Security Operations Center), where isolated AI agents work together through the Agent2Agent (A2A) protocol alongside the Model Context Protocol (MCP). This layered approach creates both vertical and horizontal security operations, allowing AI systems to share intelligence and coordinate responses to threats in real-time.

Meanwhile, CoreWeave has shared five lessons from building a multi-plane network fabric for agentic AI inference, highlighting that the infrastructure requirements for agentic workloads differ significantly from foundation model training. The company also partnered with Anyscale to caption 600 TB of video in just 95 minutes, demonstrating that hyperscale security and AI operations are becoming increasingly intertwined.

## Cloud Security and Resilience: Beyond Traditional Backup

Commvault has been making waves in the cloud security space with their comprehensive approach to backup and recovery. Their research emphasizes that ransomware now targets the entire data lifecycle, meaning that protection, detection, and recovery must be unified into a single strategy. Cloud backup and recovery solutions need to establish clean recovery points and ensure that restored operations aren't simply re-infected by dormant malware.

The company's "Ready. Or Not." podcast series finale delivered a crucial message about AI data strategies: preparation and governance matter more than raw data volume. This connects directly to the broader theme of AI security - the quality and integrity of the data feeding AI systems is as important as the security controls protecting the infrastructure itself.

## Conclusion

This week's cybersecurity developments paint a picture of an industry in transition. The explosion of AI-enabled social engineering, the emergence of new authentication vulnerabilities, and the persistent dominance of phishing as an initial attack vector all point to the same conclusion: security must evolve at the speed of the threats we face.

For security researchers and ethical hackers, the path forward involves understanding not just the technical vulnerabilities but also the human factors that make social engineering so effective. Whether it's the 7.6 billion phishing emails flowing through Microsoft's systems or the new passkey attack families unveiled at Black Hat, the underlying lesson remains the same - comprehensive security requires continuous vigilance, innovative thinking, and a willingness to challenge established assumptions about what constitutes adequate protection.

The organizations succeeding in this environment are those embracing proactive defense strategies, investing in AI-enabled security operations, and recognizing that cybersecurity is no longer just about building walls, but about creating resilient systems capable of detecting, responding to, and recovering from attacks that will inevitably come.