All items
All tracked items across vulnerabilities, news, research, incidents, and regulatory updates.
All tracked items across vulnerabilities, news, research, incidents, and regulatory updates.
12,081 items
Apple's annual financial filing warns that AI agents may take unintended or unpredictable actions, including interacting with systems they are not authorized to access, executing or altering code, sending communications and completing transactions. The company says such behavior could lead to lawsuits or harm its reputation, and that its ability to oversee third-party AI agents may be limited.
CVE-2026-115413 affects the AWS API MCP Server in versions >=1.3.0 and <1.6.0 when it runs the streamable-http transport with OAuth authentication. The server checks a bearer token's signature, issuer and expiry but not the application audience claim, so a token issued to another application in the same tenant is accepted. This may allow AWS API calls using the credentials configured on the host.
Fix: Fixed in awslabs.aws-api-mcp-server version 1.6.0. Operators using AUTH_TYPE=oauth should also set AUTH_AUDIENCE, or AUTH_CLIENT_ID for identity providers whose access tokens carry no audience claim, such as Amazon Cognito. Workaround: use the default stdio transport rather than streamable-http.
Google's Threat Intelligence Group and Mandiant report that LLMjacking attacks, in which attackers abuse stolen LLM API keys, are generating fraudulent cloud charges, with some organizations losing over $100,000 in a single day and cumulative charges reaching $1,000,000. Attackers resell access through LLM mixers such as Jet Night, which passed nearly one million dollars in token costs to one victim organization. The researchers also observed attackers exploiting CVE-2026-40217 in LiteLLM to run arbitrary code on the server.
Meta is seeking to raise between $20 billion and $25 billion in a bond sale, two-thirds in U.S. dollars and one-third in euros, to fund its spending on AI infrastructure, models and services. The sale follows S&P Global Ratings revising Meta's outlook from AA- stable to AA- negative, citing uncertainty about its capex and negative free cash flow of $2.12 billion last quarter.
OpenAI says ChatGPT now handles search queries at 20 percent of Google's volume, according to consumer product lead Adam Fry at a Monday press event. Fry defined a search as any query that triggers a web search, excluding tasks like defining a term or a simple how-to explanation. He did not specify the source of the Google volume figure.
CrowdStrike will buy Fiddler AI, which makes observability software for monitoring AI agents, and Dreadnode, which automates security testing aimed at flaws in AI models, apps and agents. According to The Wall Street Journal, the deals cost about $400 million for Fiddler AI and about $95 million for Dreadnode.
The article argues that 2025 and 2026 supply chain attacks entered widely used packages through their release processes, which ran on GitHub Actions, rather than through typosquats or malicious new packages. It lists incidents including tj-actions/changed-files (March 2025), Nx s1ngularity (August 2025), trivy-action (March 2026), and LiteLLM 1.107.0 (October 2026). The post is part one of a series, covering outside threats such as pull requests, third-party actions and downloaded scripts.
Fix: Pin every action to a full commit SHA rather than a tag or branch, because Git tags are movable pointers, and adopt a workflow lockfile. Pinned-by-SHA repositories were unaffected in the tj-actions/changed-files incident, while tag-pinned ones executed the malicious code.
An authenticated user with Machine Learning read access in one Kibana space can read audit and error messages of data frame analytics jobs in other spaces they are not authorized to access. The flaw is an authorization bypass through a user-controlled key (CWE-639). Exposed messages can include destination index names, analyzed field names and failure information.
Google describes Antigravity equipped with its new Cheminformatics Skills, which it says let an agent search thousands of compounds, run matched molecular pair analysis and propose novel rigidified macrocycles with synthetic routes in about an hour. The skills are grounded in RDKit and connect to ChEMBL and PubChem, with the language model handling orchestration while deterministic tools perform molecular calculations. The article also walks through rigidification, using crizotinib and lorlatinib as examples of adding bonds to reduce conformational flexibility and improve potency.
Google is completely turning off Search for schools that reject its AI features, according to The New York Times. In a broader education update, Google said Gemini, Gemini Notebook, Gemini in Classroom and Google Search "will be entirely off" for K-12 student accounts unless administrators explicitly turn them on. Administrators cannot disable Gemini features without also disabling Search, as Google's vice president of education, Chris Phillips, explained.
Gartner named Microsoft a Leader in the inaugural 2026 Magic Quadrant for AI Agent Management Platforms. Microsoft attributes the recognition to Agent 365, which it says began with identity through Microsoft Entra Agent ID and extends existing Microsoft Entra, Defender, Purview, Intune and Microsoft 365 admin center tools. More than 50 million agents have been registered in Agent 365.
A blog post reports that Qwen 3.8 27B, running locally on a laptop, can interpret raw neural network weights in the way the paper Sparks of Weight-Space Programming reports for GPT-6 Astra. The author tests the paper's Breakout weight-identification prompt, which is published in the paper's GitHub repository, and plans to check how Qwen 3.8 27B performs on it.
The 2026 midterm elections will shape the scope of President Donald Trump's power in his final two years and signal voter priorities ahead of the 2028 presidential election. Tech-relevant voter concerns include data centers and energy costs, AI safety, Flock surveillance expansion, and taxing billionaires' wealth in California.
Google's Threat Intelligence Group and Mandiant warned that attackers are conducting LLMjacking, a resource hijacking tactic in which they obtain an organization's LLM API key or cloud identity token through phishing, source code leaks, misconfigured environments, or exposed client-side code. Some organizations reported fraudulent charges above $100,000 in a single day, with cumulative losses reaching $1 million before access was revoked. The charges are paid by the legitimate account owner.
Cybercriminals have built a business reselling access to cloud-hosted LLMs. Some victims were charged as much as $1 million in resource consumption before access was revoked.
OpenAI has begun using the term "compute campuses" for its data centers, announcing it on Monday alongside a new "Grid Protectors Program." The program starts with an upcoming site in Effingham County, Georgia. The source frames the rebrand as an attempt to ease political backlash against AI infrastructure.
Gartner's 2026 report The Roles Required for the AI-Enabled Security Operations Center (SOC), by analyst Pete Shoard, predicts that by 2028 SOCs will have 50% more engineers than analysts as AI agents take over routine alert handling. The report says AI SOC agents already cut alert-handling work significantly for common and traditional detection use cases, and it argues analysts will shift from triage to end-to-end case ownership. Prophet Security's 2026 survey of more than 250 respondents found 57% expect team size to stay the same and 9% expect it to grow.
The source text is a newsletter navigation and teaser page listing several unrelated headlines, including a reference to the Binance hack. It contains no article content about the hack itself, so no summary can be drawn from it.
This paper studies how to suppress misaligned behaviors in language models and how durable that suppression is. The authors build two model organisms, one with a narrow trigger and one with a broad trigger, and compare SFT, GRPO and gradient-ascent unlearning on immediate suppression, durability under further training, and general capability. They introduce PEARL, a GRPO variant that augments rollouts with continuations from intermediate states along cached organism trajectories, which achieves strong suppression while preserving task accuracy across Qwen3-4B and GPT-OSS-20B.
Fix: Google recommends prioritizing Identity and Access Management (IAM) and securing the supply chain, using centralized secrets platforms to rotate and manage sensitive keys and secrets, limiting user privileges to what the role requires and denying all permissions not explicitly required, including access to Cloud LLM services, enforcing phishing-resistant multi-factor authentication (MFA), and ensuring monitoring tools give comprehensive visibility to identify unauthorized LLM and AI services (shadow AI). The source text is cut off before it finishes describing the recommendation to set API usage and spending limits.
BleepingComputerFoundation-model visual place recognition models such as SuperVLAD and BoQ keep at most 0.2% top-1 recall under a white-box ranking attack at a 1/255 budget. The proposed anchored rank-adversarial fine-tuning (ARAFT) raises robust R@1 from zero to 47.3 on SuperVLAD and 55.3 on BoQ within five clean R@1 points, where plain adversarial training yields no admissible checkpoint. ARAFT also transfers from SuperVLAD to BoQ without re-tuning and holds against stronger optimisers.