cybr.cx Daily Digest — September 02, 2026

Critical Vulnerabilities

⚠️ Actively exploited — CVE-2026-82078 & CVE-2026-81578 | PaperCut NG/MF | CVSS: Not yet scored
These two PaperCut vulnerabilities are being chained in active attacks right now — CISA added both to the KEV catalogue on August 31 with a remediation deadline of September 14. CVE-2026-81578 allows an unauthenticated remote attacker to modify system configurations (missing authentication on a critical function), which then serves as the entry point to trigger CVE-2026-82078, an unsafe reflection flaw that enables arbitrary Java bytecode execution under the PaperCut server process context. If you're running PaperCut NG or MF in any environment, patch or isolate immediately — this chain is a full remote takeover.

⚠️ Actively exploited — CVE-2026-53362 | Linux Kernel | CVSS: Not yet scored
An unspecified privilege escalation vulnerability in the Linux kernel's IPv6 networking subsystem is actively being exploited in the wild, affecting SUSE, Red Hat, and any distribution shipping a vulnerable kernel build. CISA's remediation deadline has already passed (August 30), meaning any unpatched system is overdue. Check your distribution's security advisories for patched kernel versions and prioritise reboot cycles.

⚠️ Actively exploited — CVE-2026-66384 | JFrog Artifactory | CVSS: Not yet scored
Authenticated attackers can exploit a path traversal flaw in JFrog Artifactory to write data outside the intended Docker cache directory under specific remote-repository configurations. This is in active exploitation with a CISA deadline of September 10 — teams running Artifactory as part of their CI/CD pipeline should treat this as a high-priority patch given the potential to poison build artifacts or pivot within the pipeline.

⚠️ Actively exploited — CVE-2026-60004 | Gitea | CVSS: Not yet scored
Any user with repository write access can send a malicious patch to Gitea's diffpatch API endpoint, planting an executable Git hook that runs shell commands as the Gitea service account. CISA's remediation deadline was August 28 — this is overdue. Any self-hosted Gitea instance exposed internally or externally should be patched immediately; the blast radius extends to everything the service account can reach.

⚠️ Actively exploited — CVE-2023-49105 | ownCloud | CVSS: Not yet scored
A long-standing improper authentication flaw in ownCloud allows unauthenticated file access, modification, or deletion if a target username is known and the victim has no signing-key configured. This was re-added to KEV on August 27 with a now-expired deadline, indicating fresh exploitation activity. If you're still running unpatched ownCloud, assume you've been targeted.

CVE-2026-82882 | Devtron ≤ 2.2.0 | CVSS: 8.8
Any authenticated user — regardless of privilege level — can query the /orchestrator/api-token/webhook endpoint with arbitrary parameters to retrieve plaintext super-admin JWT tokens, granting full platform control. In Kubernetes environments where Devtron is deployed, this effectively means any compromised low-privilege account becomes a cluster admin. Upgrade beyond 2.2.0 immediately.

CVE-2026-19806 | Support Genix WordPress Plugin ≤ 1.4.52 | CVSS: 8.8
A weak AES-256-CBC key derivation in the guest_ticket_login() function allows attackers to forge authentication tokens via the p parameter, leading to full administrator account takeover. The root cause is a key derived from only a handful of two-digit values — trivially brute-forced. WordPress administrators running this plugin should update or deactivate it now.

CVE-2026-8712 | Wyoming ≤ 1.10.1 | CVSS: 8.3
This Home Assistant voice processing component exposes an unauthenticated SSRF vulnerability: attackers with network access can supply arbitrary tcp:// or unix:// URIs to /api/info, /api/speech-to-text, and /api/text-to-speech endpoints to force outbound connections to attacker-controlled targets. Given Wyoming's typical deployment on home networks alongside sensitive automation, lateral movement potential is real. Upgrade to 1.10.2.

CVE-2026-84233 | RPM | CVSS: 7.0
A locally-supplied .gem file with a maliciously crafted filename containing RPM macro syntax will trigger arbitrary command execution when rpmuncompress -x processes it — the macro expands during command construction. Particularly dangerous in automated build and packaging pipelines that ingest untrusted gem files.

Headline News

Session-Stealing Malware Targets Claude AI Accounts

Threat actors are actively distributing malware designed specifically to hijack authenticated session tokens for Anthropic's Claude platform, bypassing passwords and multi-factor authentication entirely. The attack follows a well-established infostealer playbook — compromise the endpoint, harvest browser session cookies, and replay them against the AI service — but the specific targeting of Claude accounts signals that adversaries see value in monetising or weaponising access to frontier AI systems. For practitioners, this is a reminder that MFA provides no protection once a session token is in attacker hands, and that AI platforms with access to organisational data, code, and conversation history represent a meaningful new category of credential worth protecting. Endpoint hygiene, session token lifetime controls, and anomalous login detection on AI tooling should be on the checklist.

$74 Million Drained from Tectonic Protocol in Price Manipulation Attack

The Cronos blockchain was forced to halt and restart trading after a sophisticated price-manipulation exploit against the Tectonic lending protocol resulted in approximately $74 million being drained. The attack followed the classic DeFi oracle manipulation pattern: an attacker artificially moved asset prices to create a borrowing position that the protocol's logic treated as legitimate, extracting funds before the market could correct. For security practitioners watching the DeFi space, the incident underscores that smart contract audits alone are insufficient — the security of a protocol is only as strong as the integrity of its external price feeds. Cronos's network restart also raises questions about the decentralisation assurances of blockchain platforms when coordinated intervention becomes necessary to contain losses.

Honeypot Catches Coding Agents Connecting to Malicious "Free" LLM Infrastructure

A researcher running an internet-exposed inference honeypot discovered it had been silently incorporated into infrastructure offering "free" LLM API backends — relabelled with desirable model names and served to unsuspecting users and automated coding agents. Real coding-agent sessions were observed connecting to the honeypot, meaning those agents' full context windows, prompts, code, and potentially secrets were being transmitted to an adversary-controlled endpoint. The attack vector here is subtle but significant: developers and teams piping coding agents through unvetted third-party LLM endpoints may be exfiltrating codebases, API keys, and internal tooling details without realising it. Practitioners should treat LLM endpoint selection with the same supply-chain scrutiny applied to any third-party dependency.

Schrödinger's Feed

Arqit, Es'hailSat, and AIEE have demonstrated quantum-safe satellite communications in a live environment, using post-quantum cryptography to protect data transmitted via satellite infrastructure — a channel that is particularly attractive to harvest-now-decrypt-later adversaries given the multi-decade lifespan of orbital assets. Satellite communications are notoriously difficult to retrofit with new cryptographic standards, making early PQC integration at the infrastructure layer genuinely meaningful rather than just a press-release exercise. The demonstration signals that PQC is moving beyond enterprise software and into physical communications infrastructure with long replacement cycles. Practitioners responsible for critical infrastructure or government systems that rely on satellite links should watch this space — the window to influence procurement and design decisions is now, not after satellites are in orbit.

/dev/random

A developer trained a small custom transformer model in 90 minutes on consumer hardware and benchmarked it against the ARC-1 reasoning challenge — a test specifically designed to resist pattern-matching and require genuine generalisation — achieving scores that outperform several commercially deployed LLMs. The trick wasn't exotic architecture: it was aggressive dataset curation and a training objective tightly matched to the task structure, suggesting that many large models are over-parameterised for specific reasoning benchmarks rather than actually better at reasoning. It's a useful data point for anyone who's been told they need a hundred-billion-parameter model to solve their problem. Sometimes 90 minutes and a clear objective beats a data centre.