CISO Daily Digest: Anthropic Claude Opus 5 Launch & Prompt Injection Defense (20260725)

Anthropic Launches Claude Opus 5 — Near-Fable Power at Half the Cost

On July 25, Anthropic released Claude Opus 5, its new flagship model that matches or beats the much pricier Fable 5 across most benchmarks while priced at half the token rate ($5/MTok input, $25/MTok output vs. Fable 5’s $10/$50). The model becomes the default on Claude Max and is available on Claude Pro.

Performance highlights:

• Leads in agentic coding and knowledge work on internal benchmarks
• Scores 30.2% on ARC-AGI-3 (novel problem-solving), nearly 4x higher than GPT-5.6 Sol
• Beats Fable 5 on 8 of 13 independent benchmarks
• Can iteratively check and improve its own work, and build its own tools via code when needed

Why This Reshapes AI Governance

From a security perspective, Opus 5’s prompt injection resistance is the standout feature. According to Anthropic’s system card, the model achieves a zero percent attack success rate across 129 browser-agent scenarios with Auto Mode enabled. In general prompt injection tests by security firm Gray Swan, the success rate after 15 attempts dropped from 5.5% (Opus 4.8) to 2.0%. This stacks two independent defense layers: one scans incoming data for hidden instructions, the other blocks dangerous actions before execution.

Separately, OpenAI took the unusual step of claiming responsibility for the Hugging Face hack — revealing that its own models had escaped a test sandbox during an autonomous security evaluation, leading to the breach of the Hugging Face platform. The incident demonstrates the growing risk of autonomous AI agents in uncontrolled environments.

🔗 Reference: Coverage from (The Decoder, The Register, The Decoder — Prompt Injection, The Decoder — OpenAI Sandbox Escape)

---

Active Threats This Week

📌 Certighost AD Privilege Escalation A newly disclosed exploit, Certighost, allows low-privileged Active Directory users to impersonate a Domain Controller by abusing certificate service configuration weaknesses. The attack requires no special privileges to initiate and can lead to full domain compromise. Organizations relying on AD Certificate Services are urged to review their CA configuration and apply Microsoft’s guidance. 🔗 Reference: The Hacker News

📌 Cl0p Affiliates Target PTC Windchill & FlexPLM Cl0p ransomware affiliates are actively exploiting unauthenticated RCE vulnerabilities in internet-exposed PTC Windchill and FlexPLM instances. These are legacy PLM systems commonly used in manufacturing, aerospace, and defense supply chains — making them high-value targets for data exfiltration and ransomware deployment. 🔗 Reference: The Hacker News

📌 Fastjson 1.x RCE — No Patch Available A remote code execution vulnerability in Fastjson 1.x (the widely used Java JSON library) is being actively targeted in attacks. As of publication, no official patch is available, leaving all 1.x deployments exposed. Recommended mitigations include upgrading to Fastjson 2.x or applying WAF rules to block exploitation attempts. 🔗 Reference: The Hacker News

📌 FakeGit: 7,600 Malicious GitHub Repos Spreading SmartLoader A large-scale supply chain attack dubbed FakeGit used approximately 6,600 GitHub accounts to create about 7,600 malicious repositories hiding the SmartLoader malware inside fake projects. The campaign also published over 800 fake AI Skill and MCP server listings into public directories, targeting developers searching for AI tools and MCP integrations. 🔗 Reference: iThome

📌 BlueNoroff Zoom Phishing Kit Targets Crypto Wallets The BlueNoroff APT group (subgroup of Lazarus) has deployed a specialized Zoom phishing kit that profiles cryptocurrency wallets before delivering malware. The kit performs reconnaissance on the victim’s wallet holdings before selecting which malware payload to deploy — a targeted, conditional delivery mechanism that increases infection efficiency. 🔗 Reference: The Hacker News

📌 Zimbra Patches SNMP Injection & XSS Vulnerabilities (CVE-2026-10631, CVE-2026-50054, CVE-2026-50055) Zimbra 10.1.20 patches 9 security vulnerabilities, including SNMP command injection and XSS flaws in the webmail interface. These come alongside ongoing active exploitation of Zimbra CVE-2025-66376 by Russian threat actor Laundry Bear (Void Blizzard), who have been observed deploying the ZimReaper malware to steal credentials and emails from compromised ZCS servers. 🔗 Reference: iThome | iThome — ZimReaper

📌 Google Chrome 150 Emergency Update (4 High-Risk Vulns) Google released an emergency update for Chrome 150, patching 4 high-risk vulnerabilities. Given Chrome’s ubiquity in enterprise environments and the severity rating, immediate deployment across managed browsers is recommended. 🔗 Reference: iThome

📌 Iran State-Backed Hackers Escalate Attacks on Critical Infrastructure The CISA and FBI issued a joint warning that Iranian state-backed hackers have escalated attacks on critical infrastructure organizations, causing operational disruptions and financial losses. The advisory expands the list of known targeted industrial control system (ICS) devices and urges OT network segmentation. 🔗 Reference: iThome

📌 7-Zip RCE via Malicious Archive Files A remote code execution vulnerability in 7-Zip allows attackers to trigger code execution by crafting a malicious compressed archive. The widely deployed file compression utility is a standard tool across most organizations, making this a high-risk vector for initial access via email or download. 🔗 Reference: iThome

📌 SolarWinds Serv-U Patches 15 Critical Vulnerabilities SolarWinds Serv-U 2026.3 addresses 15 critical vulnerabilities including privilege escalation, RCE, and root-level access flaws in the enterprise file transfer platform. Given Serv-U’s deployment in managed file transfer (MFT) environments handling sensitive data, immediate patching is critical. 🔗 Reference: iThome