Threats Tagged 'vulnerability'

A critical vulnerability (CVE-2026-20230) in Cisco Unified Communications Manager (Unified CM) has been recently patched but is currently being exploited in attacks. The flaw allows unauthenticated remote attackers to perform SSRF attacks, write arbitrary files to the operating system, and escalate privileges to root, but exploitation requires the WebDialer service to be enabled, which is disabled by default. Proof-of-concept code was publicly available at the time of patch release, and exploit intelligence firm Defused observed active exploitation from a single source. Cisco has not yet confirmed in-the-wild exploitation. Unified CM is a core enterprise communications platform, making this vulnerability potentially valuable to attackers. No affected versions were explicitly stated in the source information.

This vulnerability concerns a potential alignment issue in large language models (LLMs), specifically observed in Gemma-3-12B and initially noted in Claude. The issue arises when a long, structured text influences the model's internal hidden states, causing measurable changes in how the model processes subsequent unrelated tasks. This effect can temporarily weaken the model's alignment constraints, making it more willing to provide direct or controversial answers that it would otherwise avoid. The phenomenon is behavioral and supported by hidden-state analysis in open-weight models, but no permanent model modification or universal alignment bypass has been demonstrated.

A Broken Access Control vulnerability was discovered and responsibly disclosed in a government student welfare portal used by over 300,000 students in India. The flaw allowed unauthorized authenticated users to access privileged functionality and sensitive beneficiary information, including addresses and government benefit details. The vulnerability stemmed from authorization being enforced only on the frontend, without proper backend validation. The issue was reported to CERT-In and relevant authorities, and has since been confirmed fixed.

A critical vulnerability (CVE-2026-20253) in Splunk Enterprise allows unauthenticated attackers to create or truncate arbitrary files via a PostgreSQL sidecar service endpoint lacking authentication controls. The flaw affects Splunk Enterprise versions 10.2 before 10.2.4 and 10.0 before 10.0.7. Exploitation was confirmed shortly after public disclosure, with proof-of-concept code published. CISA has added this vulnerability to its Known Exploited Vulnerabilities catalog and mandated rapid patching for federal agencies. Splunk has released patches to remediate the issue and strongly recommends upgrading to fixed versions.

This entry references a vulnerability summary bulletin published by CISA for the week of June 8, 2026. The information is a pointer to an external source without specific vulnerability details included in the input. No affected software versions or detailed technical descriptions are provided.

A chain of three vulnerabilities in LiteLLM, an open-source AI gateway, allows a low-privilege user to escalate to full admin and execute arbitrary code on the server. The vulnerabilities include an authorization bypass, privilege escalation, and a sandbox escape leading to server-side code execution. Exploiting the chain exposes all provider keys, stored credentials, and data passing through the gateway, enabling an attacker to alter AI model responses and potentially execute arbitrary commands on connected systems. The maintainer released a full fix in LiteLLM v1.83.14-stable.

A high-severity vulnerability (CVE-2026-5027) in the Langflow low-code AI development platform allows unauthenticated attackers to write files to arbitrary locations via a path traversal flaw in the 'POST /api/v2/files' endpoint. This flaw enables remote code execution (RCE) because the filename parameter is not sanitized, and Langflow's default unauthenticated auto-login allows attackers to reach the vulnerable endpoint without credentials. Exploitation attempts have been observed in the wild, with attackers dropping test files on victim systems. Approximately 7,000 Langflow instances are internet-accessible, mostly in North America. The vulnerability was publicly disclosed in March 2026, and no patch or official fix information is provided in the source content.

A vulnerability referred to as the 'GreatXML bitlocker bypass vulnerability' has been publicly disclosed via a Reddit post linking to an external Git repository. The information is minimal, with no detailed technical description, affected versions, or vendor advisories provided. No known exploits in the wild have been reported. The severity is assessed as medium based on the limited available data.

This entry discusses research on using large language models (LLMs) to assist in vulnerability discovery. The research involved comparing manual vulnerability discovery in Zenitel's TCIV-3+ video intercom with automated analysis using Anthropic's Claude Opus 4.6. The findings highlight nuanced results about the capabilities and limitations of AI-driven vulnerability research. No specific new vulnerability details, exploits, or affected versions are provided. There is no indication of active exploitation or patches related to this research. The post primarily serves as an exploration of AI's role in vulnerability research rather than reporting a direct security threat.