All items
All tracked items across vulnerabilities, news, research, incidents, and regulatory updates.
All tracked items across vulnerabilities, news, research, incidents, and regulatory updates.
12,118 items
Joe Security has launched Adaptive Detonation, a feature of Joe Sandbox AI that uses AI agents to find obstacles stopping a malware sample from showing its behavior, such as timing checks or missing dependencies. The agents then apply a targeted intervention, like patching a blocking check or supplying missing command-line arguments, and re-run the sample to verify whether more activity appears. The plugin is available for Joe Sandbox Cloud.
OpenAI described a five-stage agentic pipeline, built around an upcoming cyber-focused model, that found and validated Linux kernel flaws. In a mid-October 2026 run lasting about 30 hours, it reproduced 1,724 vulnerabilities and generated 948 working local privilege escalation exploits that could let an unprivileged local user gain root access, at a total API cost of $1,213,430.
A vulnerability in myshell-ai AIlice up to 0.4.0 affects the function LoadModel_PEFT in ailice/core/llm/AModelCausalLM.py, within the HuggingFace Model Wrapper component. Manipulating the argument base_model_name_or_path leads to inclusion of functionality from an untrusted control sphere, and the attack can be launched remotely. The exploit is public, and the project was informed through an issue report but has not responded.
A path traversal flaw exists in the Model Path Resolution component of Edge0-AI Audio8_TTS, up to commit 1d64afcc2da56222b41ddcf53de7b8bc54be1fb6, tracked as CVE-2026-122159. It sits in ArkTtsRuntime.__init__ in onnx_runtime_0_1b_int8/arktts_runtime/runtime.py, where manipulating the slow_decode_models argument can lead to path traversal. The attack can be carried out remotely, and a public exploit is available.
Open edX Platform, from the Ulmo release until release/ulmo.4 and release/verawood.1, exposes GET /api/support/v1/manage_course_team/ in CourseTeamManageAPIView with only IsAuthenticated, omitting the global support-role check. Any authenticated learner can query arbitrary targets by email, username or user_id and tell active accounts from absent or inactive ones through the 200 and 404 response difference. Callers with a course-level or organization-level instructor role can also read the target's staff or instructor assignments within their own scope.
Victoria's major parties are trading attacks over AI-generated videos of party leaders in election advertising. The source describes a Halloween-night social media post in which Victorian Liberal leader Jess Wilson warns of "Americanised healthcare" and claims a heart attack could cost $150,000 in care. Transparency and advertising experts quoted in the piece say such "lies, deepfakes and AI-generated politicians" could further erode trust in politics.
Jan, from version 0.8.5 before 0.8.8, contains a workspace guard bypass in the agent-tools gate.rs file. Attackers who supply a malicious repository can use prompt injection to make the write or edit tools follow dangling symlinks, letting them create files outside the workspace, such as ~/.ssh/authorized_keys on the host.
Mistral Vibe before 3.0.3 contains a remote code execution vulnerability in project context collection. A crafted repository that sets gpg.program and log.showSignature, combined with a signed commit, causes git log to run the attacker's program at session start, before any trust prompt.
Fix: Fixed in 3.0.3 (the source states Mistral Vibe before 3.0.3 is affected; the upgrade to 3.0.3 or later is implied by that version boundary).
Mistral Vibe before 2.25.7 contains a remote code execution flaw, tracked as CVE-2026-122134. An attacker can configure malicious Git clean or process filter drivers in a repository, and when a user opens that crafted repository, ProjectContextProvider._fetch_git_status runs git status at startup. This invokes the filters and executes the attacker's commands with user privileges before any trust prompt is shown.
Mistral Vibe before 2.25.5 on Windows contains an untrusted search path flaw. An attacker can place a malicious git.exe in a repository root, and SessionLogger and ProjectContextProvider run it at session start before any trust prompt, so luring a user into opening a crafted repository is enough to execute code.
Fix: Fixed in 2.25.5.
NVD/CVE DatabaseAT&T's Connected Life smart home security system now integrates with Gemini for Home. AT&T says it is the first home security product to integrate with Google Home Premium. The integration adds Gemini-powered features that summarize camera events and let users search recorded video using natural language.
Hugging Face CEO Clément Delangue said at the WIRED World Fair in Miami that most future AI risk will come from a few companies building behind closed doors, and that open ecosystems will mitigate it. He said his P(doom) is "close to zero" and argued that those who deploy AI agents should be legally responsible for their actions, including agentic cyberattacks.
Endor Labs explains core terms in agentic development: models, harnesses, hooks, skills, plugins, and MCP servers. It describes where each sits in the agentic development lifecycle and how each creates an AppSec entry point or attack surface. The post cites IDC's July 2026 DevSecOps and Application Security Survey, in which 79.6% of respondents reported at least a moderate reassessment of their AppSec tools due to code security capabilities from frontier AI companies.
TechCrunch's Equity podcast interviewed Irene Solaiman, Chief Policy Officer and SVP of Strategy at Hugging Face, about building on open source AI models. The discussion covers cost and control motivations for switching to open models, how companies capture value when models are free, and whether the U.S. can compete with Chinese open models. It also addresses how to keep an open AI ecosystem secure after recent attacks on open AI platforms.
ChatGPT can now create animated GIFs from a text prompt or an uploaded photo. The feature rolls out alongside OpenAI's Images 2.6 model and lets users adjust animation speed, edit individual frames, and share the result. Previously, ChatGPT could not generate GIFs natively in a single step.
VicOne and Mobile Defender Inc. published a joint blog on how GPS spoofing lets organized thieves feed fleet tracking systems false location data during cargo theft. The post describes a three-layer approach to detect, correlate, and respond to location manipulation. It draws on Verisk CargoNet figures showing Q2 2026 losses rose to $304.6 million even as incidents fell 26 percent.
Fix: The source describes a three-layer approach (detection at the vehicle, correlation across the fleet, and a structured response playbook) but the excerpt is truncated before stating specific mitigation steps.
VicOne BlogCheck Point Research disclosed a jailbreak bundle named Haiou, which modifies OpenAI Codex configuration files to insert advertisements for a third-party OpenAI-compatible LLM gateway into agent responses. The bundle is distributed via a GitHub repository called haiou-codex and bundled within the open-source tool Codex-X, and Check Point codenamed the behavior prompt adware. Researchers warn that similar third-party skills or configurations could be used for supply chain attacks.
Endor Labs was named a Leader in The Forrester Wave: Agentic Development Security Tools, Q4 2026, and received the highest possible score in seven criteria, including innovation, AI application flaws, AI SCA, and AI coding toolchain. The article argues that agentic development security must cover coding agents, the packages they install, and their workflows, not just the code they write. It also cites a Concordia University study finding that automated code review consumed 59.4% of tokens on average across 30 tasks.
Uber executives disclosed on an earnings call that third-party AI agents arranged only four rides in the third quarter of 2026. The company also recorded around 500 Uber Eats orders fulfilled by AI agents during the week of Halloween, out of 320 million total rides and deliveries. This is the first time Uber has disclosed such figures.
Fix: Fixed in release/ulmo.4 and release/verawood.1.
SecureGARD is a dynamic ontology-guided governance framework for tool-using LLM agents. It maps heterogeneous tool calls to a shared semantic representation and uses an executable ontology with episode-level task state to issue ready, waiting, or forbidden decisions. Evaluated on AgentDojo, ASB, and AgentDyn under benign and indirect prompt-injection settings, it reports a 0% attack success rate across all three benchmarks while improving task utility under attack.