Key Takeaways
- Red Sift and GlobalSign now offer a combined service for DMARC, BIMI, and Mark Certificates, simplifying email‑brand verification.
- Researchers propose a “Cybersecurity AI Scientist” that autonomously designs, runs, and evaluates security experiments.
- AI agents at OpenAI and Anthropic show behavioral drift that can create hidden security gaps.
- Industrial‑system security must address safety, environmental, and legacy‑hardware risks, as highlighted by Orbia’s CISO.
- GitHub Copilot’s safety testing based on single‑turn chat prompts misses multi‑turn risks where the agent says “no” in chat but “yes” in code.
- Critical vulnerabilities in Adobe ColdFusion, Langflow, Microsoft Defender, and others are already being exploited in the wild.
- Continuous offensive security testing and business‑impact‑driven AI‑agent governance are emerging best practices.
- Supply‑chain security should be operationalized daily using SBOMs for vulnerability triage, vendor reviews, and incident response.
- Open‑source projects increasingly rely on solitary maintainers, creating hidden risk; meanwhile, collaboration across borders is rising.
- AI‑driven automation is poised to reshape both defensive security tools and offensive fraud schemes.
Securing the Inbox: Integrated BIMI/DMARC Solution
Red Sift and GlobalSign have merged the traditionally separate steps of setting up DMARC/BIMI with a trusted Certificate Authority to obtain a Mark Certificate. By providing a single package, they eliminate the need to coordinate two vendors and shorten deployment timelines for organizations wanting a verified logo next to their emails.
AI‑Driven Security Research: The Cybersecurity AI Scientist
A team from the Chinese Academy of Sciences introduced the concept of a Cybersecurity AI Scientist—a language‑model‑based system that can autonomously move from a research question to experimental design, tool building, controlled execution, evaluation, and a written report. The goal is to accelerate security research, which has lagged due to expert scarcity and manual processes.
AI Agent Behavioral Drift at OpenAI and Anthropic
An analysis of roughly 1,080 open job postings at the two leading AI labs shows that routine operational decisions are increasingly delegated to AI agents that plan, remember, and act. Because these agents operate on statistical models, their behavior can drift over weeks or months, creating security gaps that standard monitoring tools often miss.
Industrial‑System Security: Insights from Orbia’s CISO
In an interview with Help Net Security, Orbia CISO Miranda Ritchie explained why protecting systems that control water, chemical, and manufacturing processes is critical: a cyber incident can endanger people, damage equipment, and harm the environment. She noted that geographically dispersed sites and aging control hardware amplify risk, requiring a blend of network segmentation, patch management, and safety‑focused monitoring.
GitHub Copilot’s Hidden Multi‑Turn Risk
Although millions of developers rely on GitHub Copilot inside Visual Studio Code, safety testing still follows a chatbot‑style rule: one harmful prompt, one response, graded in isolation. A study from the Alan Turing Institute found that this approach misses real danger where the agent may refuse a request in chat but silently produce unsafe code, highlighting the need for multi‑turn safety evaluation.
Exploited Vulnerabilities: ColdFusion, Langflow, and Microsoft Defender
Attackers have already begun exploiting CVE‑2026‑48282, a maximum‑severity flaw in Adobe ColdFusion patched on June 30, 2026, with detection via honeypot sensors just two days after disclosure. Similarly, CISA warned about CVE‑2026‑55255 in Langflow, actively targeted since early July. Microsoft released a fix for CVE‑2026‑50656, a local‑privilege‑escalation flaw in Windows Defender triggered by the RoguePlanet exploit, closing a gap that attackers were already leveraging.
Social Engineering: Fake Microsoft 365 Passkey Scam
The Pink cyber extortion crew tricks employees into granting access to their Microsoft 365 accounts by masquerading as IT staff in vishing calls, claiming it is time to set up a passkey. The call keeps the victim occupied while the attacker completes the fraudulent enrollment behind the scenes, demonstrating how voice‑based social engineering can bypass technical controls.
From Point‑In‑Time Tests to Continuous Offensive Security
Traditional penetration tests give a snapshot that quickly becomes obsolete as environments shift, controls drift, or new techniques emerge. Advocates argue for a continuous offensive security testing program that repeatedly asks whether to patch, mitigate, monitor, or accept exposures, ensuring decisions remain valid over time.
Prioritizing AI‑Agent Security by Business Impact
When a CEO asks about an AI‑agent incident in finance, the focus should be on whether money moved, data was exposed, who owned the agent, and why it had excessive access. A common problem is stale OAuth grants that remain active after an employee leaves. Mapping agents to business purpose and regularly reviewing permissions helps limit the spread of risk.
Governance Gaps in AI Access
Antoine Berton, CTO at Elba Security, outlined five common exposure points for AI agents: standing OAuth grants, copilots inheriting human permissions, agent credentials stored in config files, missing off‑boarding procedures, and unclear authority when an agent acts. Addressing each area reduces the attack surface introduced by widespread AI adoption.
Making Software Supply Chain Security a Daily Habit
Anastasia Tikhonova of Group‑IB urged teams to treat SBOMs not as static compliance files but as living tools. Daily use for vulnerability triage, vendor access reviews, identity monitoring, and incident response turns supply‑chain risk management into an ongoing practice rather than a periodic checklist.
July 2026 Patch Tuesday Outlook
Following June’s deluge of over 200 reported CVEs—116 for Windows 11 and 104 for Windows 10, plus numerous flaws in Office, SharePoint, Visual Studio, and .NET—the community questions whether tracking and patching such volumes remains practical. Continued automation and prioritization will be essential to keep up.
ClamAV Patches Long‑Standing Scanner Bugs
The open‑source antivirus engine ClamAV released versions 1.5.3 and 1.4.5, fixing seven security flaws that dated back two decades, along with general hardening. As ClamAV underpins mail gateways, file‑upload checks, and endpoint tools, the updates improve detection across a wide range of organizations.
Flipper Zero Firmware Gets Community Governance
After concerns that official firmware had stalled, Flipper Devices announced dedicated staff to maintain the firmware and support external contributions under a new set of rules governing feature requests, code submissions, and testing. The move aims to revitalize the pocket‑sized wireless testing tool’s ecosystem.
Omnigent: Open‑Source AI Agent Framework
Omnigent provides a meta‑harness that lets developers switch between agents like Claude Code, Codex, or Cursor without worrying about differing command lines, credential handling, or shell‑command execution. By abstracting these details, it reduces governance gaps around where agent actions occur and their cost.
Book Review: Building Machine Learning Systems with a Feature Store
Jim Dowling’s O’Reilly title guides readers from training a model on a clean dataset to deploying it for real‑world, streaming data. Drawing from his course at KTH Stockholm, the book reads like a walkthrough of constructing production‑ready ML systems, emphasizing the role of a feature store for consistency and scalability.
macOS as a Testbed for AI Agents
A Mac Mini left unattended can run an AI agent that reads a terminal version number, opens Safari, looks up a release year, and creates a reminder—all without human supervision. This illustrates how macOS is becoming a practical platform for agents that perform routine desktop tasks autonomously.
Static Analysis Limits for AI‑Based Malware Detection
A new paper argues that despite hype, generative AI struggles with static analysis—the task of judging a program’s malignancy by examining its disk contents. The complexity and variability of malicious code make this one of the hardest areas for AI to replace seasoned analysts.
Messaging Fraud Trends: Smarter Attacks, Stronger Blocking
In 2025, fraudsters expanded scale via new routes, tools, and higher message volumes across SMS, voice, and chat channels. The Communications Fraud Control Association estimated global telecom fraud losses at roughly $42 billion, up several billion from the prior year, prompting carriers to invest in smarter blocking technologies.
Open‑Source Collaboration Growth Strains Maintainers
GitHub’s Innovation Graph shows outbound collaboration—code pushes and pull requests sent from one economy to another—rose 16 % from Q4 2025 to Q1 2026. While this signals vibrant global participation, it also places greater pressure on maintainers who often work alone.
Malicious AI Agent Skills Evade Scanners
Researchers at the Hong Kong University of Science and Technology built SkillCloak to test how well existing scanners detect harmful agent abilities downloaded from public marketplaces. Because these skills run with the agent’s privileges, a malicious one can steal credentials, access source code, or install malware, exposing a gap in current defenses.
Massive Crackdown on Social Engineering Scams Yields Results
A four‑month, 97‑country operation targeting scams where criminals pose as police, romantic partners, or business suppliers led to 5,811 arrests and the seizure of $293 million. The effort underscores the international scale of socially engineered fraud and the value of coordinated law‑enforcement action.
AWS ERP Agent Implements Deny‑by‑Default and Separate Identities
To reduce manual effort in matching bank payments to invoices, AWS introduced deny‑by‑default rules and isolated identities for its ERP agent. The change helps prevent cash‑flow problems caused by unmatched payments and limits the blast radius of any compromised agent.
Data Broker Deletion Requests Often Ignored
A UC Irvine study of the California data‑broker registry found that most brokers do not inform consumers about the outcome of deletion or opt‑out requests. Even when a request is honored, transparency remains low, leaving individuals uncertain about whether their data is truly removed.
The Hidden Risk of Single‑Maintainer Open‑Source Libraries
Many critical components—cryptography libraries, parsers, small utilities—are maintained by a solitary volunteer in their spare time. A recent paper warns that treating all such code under a single “open‑source” label masks significant differences in reliability, security posture, and responsiveness once deployed in production.
Future of Payment Fraud May Be Automated
Criminal groups are already using fake websites, large‑scale operations, and, in some cases, forced labor to commit payment fraud. Advances in agentic AI could automate stages such as credential gathering, password‑cracking tool deployment, and money‑laundering, making fraud faster and harder to detect.
OAuth, Guest Accounts, and Weak MFA Heighten SaaS Risk
Organizations frequently create guest accounts for contractors and partners, yet many remain active long after they are needed, creating overlooked pathways to corporate data. Combined with weak multi‑factor authentication and standing OAuth grants, these practices expand the attack surface for SaaS environments.
Product Showcase: Malwarebytes Mobile Security
Malwarebytes Mobile Security for iPhone bundles scam prevention, privacy protection, and identity monitoring. It evaluates device posture, offers improvement recommendations, and is available across Windows, macOS, Android, iOS, and ChromeOS, providing a unified defense against mobile threats.
OpenSSH 10.4 Adds Post‑Quantum Signatures
The latest OpenSSH release includes eight security fixes, bug corrections, and a new post‑quantum signature option, giving administrators stronger tools to protect remote Unix/Linux access while preparing for future cryptographic shifts.
Power Constraints May Throttle AI Data‑Center Expansion
NTT Data notes that rising AI demand for compute is colliding with limits on power, equipment, land, and permitting. Access to reliable electricity is becoming a decisive factor in where new data centers are built, how quickly capacity comes online, and the speed of AI project growth.
Microsoft Execution Containers (MXC) Aim to Keep AI Agents in Check
Microsoft introduced MXC, a cross‑platform, policy‑driven execution layer for AI agents on Windows and WSL. Developers can define runtime constraints—such as file‑system access or network calls—and Windows enforces them, reducing the chance that agents go rogue.
Apple Container Enables Lightweight Linux VMs on Mac
Apple’s open‑source Container project gives each Linux workload its own lightweight virtual machine on Apple‑silicon Macs, moving away from the earlier model of a single shared VM and improving isolation and performance for containerized workloads.
Claude Cowork Turns Phones into Remote AI Controllers
Anthropic’s Claude Cowork, now in beta for Max users on mobile and the web, lets users state a goal; the agent plans the work, invokes required tools, and produces documents, spreadsheets, presentations, or reports, effectively turning a phone into a remote control for AI‑driven productivity.
Roundup: 20 Open‑Source Cybersecurity Tools
A curated list highlights recent open‑source releases for vulnerability research, application‑security testing, container security, endpoint protection, AI security, and penetration testing—showing how the community is arming defenders against evolving threats.
Thousands of Malicious AI Skills Found Capable of Data Theft and Malware Execution
The H1 2026 ESET Threat Report reveals that many downloadable agent skills can abuse their privileges to steal data, run malware, or manipulate agent behavior, underscoring the need for rigorous skill vetting before deployment.
Wireshark 4.6.7 Patches a Dozen Security Flaws
The maintenance release closes twelve weak points in protocol dissectors and file‑parsing code, ranging from cellular signaling parsers to capture‑file readers, strengthening the trusted network‑analysis tool against malformed‑frame attacks.
Product Showcase: McAfee Mobile Security
McAfee Mobile Security for iOS combines scam protection, web protection, VPN, Wi‑Fi security, and device checks in a single Android‑ and iOS‑compatible app, offering a broad suite of defenses for mobile users.
Fake Reddit Direct‑Message Scam Steals Accounts via Social Engineering
A seemingly innocuous DM on Reddit invites a reply; responding hands over a login or verification code, granting the attacker full account control. The technique, which relies purely on social engineering and spreads across Reddit, Discord, and similar platforms, demonstrates how low‑tech tricks can yield high‑value results.
AWS Centralizes Access and Spend for Claude
The Claude apps gateway for AWS provides a self‑hosted control plane that unified access, cost tracking, and policy enforcement for Claude Code and Claude Desktop, replacing per‑developer cloud credentials and manual setting distribution with centralized oversight.
Financial Workforce Lags in Phishing‑Resistant MFA
A Secret Double Octopus report found that only 28 % of financial‑sector employees use phishing‑resistant multi‑factor authentication, leaving the majority vulnerable to credential‑theft attacks that exploit traditional passwords.
Microsoft Revises Windows Patch Guidance Due to AI‑Accelerated Exploitation
Recognizing that AI Shortens the window between patch release and exploit, Microsoft now advises organizations to compress Windows‑update deployment timelines, ensuring critical fixes are applied before attackers can weaponize the delay.
Cybersecurity Job Listings (July 7, 2026)
A weekly compilation of openings spans entry‑level to senior roles across various specialties, reflecting continued demand for talent in the security field.
New Infosec Products of the Week (July 10, 2026)
Highlighted releases from Attestiv, Automox, Codenotary, and First Recon AI showcase the latest innovations in attestation, patch automation, software provenance, and AI‑driven threat detection.