TokenPost topic
AI security
91 TokenPost stories are tagged AI security. The newest appear first.
Latest AI security stories
- AI Campaign Reached 10 Organizations, With Unauthorized Access or Data Retrieval Confirmed
Human operators directed 25 operations against at least 19 organizations in August and September 2026. The findings do not establish a success rate.
· Technology
- Microsoft Researchers Use Seven-Neuron Edit to Trigger an AI Marker
The seven-neuron test predicted a harmless command marker in recorded contexts. A separate deletion-command test measured predictions, not execution.
· Technology
- NEAR Protocol Adds Investment Thesis Track to Its Hackathon
Participants will use on-chain activity data and live dashboards to support cases for the network’s long-term prospects.
· Technology
- Cursor Says Most Customers Were Unaffected by October Security Incident
After a two-week investigation, Cursor said it notified affected customers directly and traced the attack to an access key exposed at Fireworks.
· Technology
- Ethlabs Member Says Ethereum’s pqai27 Plan Still Awaits Governance Review
Barnabé Monnot said the proposal reflects Ethlabs’ view and lacks full agreement. The group’s Quick Slots pledge applies to Hegotá, while it still plans to pursue quicker slots in 2028.
· Technology
- Island Maps 186 AI Lookalike Domains, Traces Malware Routes
The inventory spans 27 brands and includes parked domains; 186 is a domain count, not an infection tally.
· Technology
- Claroc Finds 35,578 Grok Build App Addresses Accessible Without Sign-In
After Claroc notified SpaceXSI, 24 apps created by SpaceXSI and X employees were removed. SpaceXSI said apps remain private until creators share them.
· Technology
- Ethlabs Proposes Post-Quantum and AI Security as Ethereum’s 2027 Priority
The proposal calls for reassessing the Hegotá hard fork, dropping quick slots and other noncritical features, and treating pqai-scaled as an extended goal.
· Technology
- Google DeepMind Uses AI Agents to Verify Signal’s Cryptographic Code
The project produced more than 6 million lines of Lean-checked proofs covering forward secrecy, authentication and post-compromise security.
· Technology
- Fireworks Confirms Second Breach as Large Customers Weigh Exits
The Nov. 9 disclosure covered customer billing information, datasets and unfinished model weights. Two customers have moved most of their business away from the AI provider.
· Technology
- OpenAI Restricts GPT-6 Cyber Access as Anthropic Unveils FlashProve
GPT-6 Cyber is exclusive to vetted Daybreak Red teams. FlashProve is an open-source solution for verifying AI model outputs.
· Technology
- OpenAI Partners With Hacktron on System Security
The partnership will test OpenAI’s connections to third-party services as the company works to strengthen its defenses.
· Technology
- India's CERT-In Warns Crypto Firms of Active CRAFTY FALCON Attacks
The campaign targets financial firms, abuses payment APIs and uses public AI-driven tools to tailor attacks.
· Technology
- Five Targeted Documents Cut Classical RAG Retrieval Scores by Up to 61.5%
In isolated-target tests across 623 SciFact and NFCorpus queries, five documents were tailored to each query. The results measured retrieval quality, not AI-generated answers.
· Technology
- Binance’s 何一 Says Easy Financing Weakens Incentives to Keep Working
The Binance co-CEO also discussed lower-market-cap memecoin listings, tokenized stock and commodity products, and security defenses.
· People
- Binance Plans AI Security Fund After $6.34 Billion Theft
The exchange hired outside cybersecurity experts to review the Oct. 29 incident and plans to expand monitoring, endpoint protection and developer security.
· Technology
- Mistral Says Mistral Large 4 Coordinated 133 Agents in Django Security Demo
The video ran at 90 times normal speed. Django had published fixes for six issues on Nov. 3, but Mistral did not identify the tested flaw or show its patch.
· Technology
- CrowdStrike Agrees to Buy Dreadnode to Add AI Security Testing
Dreadnode’s tools test AI systems against more than 100 known attack strategies. The acquisition is expected to close in the fourth quarter of fiscal 2027.
· Business
- 赵长鹏 Urges Web3 Industry to Expand AI Security Defenses
Web3 participants are also discussing an industry security alliance, though few details have been disclosed.
· Technology
- CrowdStrike Launches AI Emergency Switch for Abnormal Agents
The tool can terminate agents, block them or cut their network connection. CrowdStrike has also agreed to acquire Fiddler AI, whose software monitors and explains AI agent behavior.
· Technology
- CrowdStrike Agrees to Buy Fiddler AI to Expand Agent Monitoring
The planned payment is mostly cash, with a stock component subject to vesting. CrowdStrike did not disclose the purchase price.
· Business
- OpenAI Plans Security Sprint With Cathedral to Test Its Defenses
The exercise will bring together OpenAI teams, new researchers and security partners to use the company’s models in defensive testing. The announcement gave no schedule, duration or scope.
· Technology
- BTIG Lifts Netskope Price Target to $30 From $22, Keeps Buy
BTIG cited stronger demand checks, competitive gains against Zscaler and potential data-protection demand as AI security develops. Netskope's fiscal second-quarter revenue rose 29% to $220.5 million.
· Insights & Views
- CoreWeave and CrowdStrike Expand Partnership for AI Security
CrowdStrike will offer Falcon services to CoreWeave Cloud customers, while CoreWeave will provide the AI platform for CrowdStrike SafeMind.
· Business
- Reddit User Says Hugging Face Mythos-Class Weights Look Synthetic
A limited sample led Reddit user johnxreturn to estimate a 0.5% chance that the files came from an authentic Anthropic checkpoint.
· Technology
- South Korea’s Financial Regulator to Review Security With Five Bank Chiefs
The meeting is set for 8 p.m. ET Wednesday, Nov. 4 (1 a.m. UTC Thursday, Nov. 5), with AI-related threats on the agenda.
· Regulation
- NEAR Protocol Sets AI Security Track for Its Upcoming Hackathon
The track focuses on protecting digital assets and spotting scams, with policy engines and sandboxed wallets among the proposed projects.
· Technology
- Anthropic Restored Apps After Cathedral Agents Found Paths to Staff Records
Cathedral said its agents found routes to salary data, job offers, meeting transcripts and Slack messages. Anthropic said customer prompts, model responses, source code and model weights were not accessed.
· Technology
- OpenAI Details Expanded Monitoring That Flagged Limited DNS Chatbot Access
During a Sept. 20 reinforcement-learning run, the system flagged limited DNS-based access to an external chatbot within 15 minutes; a human began review three minutes later.
· Technology
- Eclypsium Demonstration Shows AI Model Weights Can Change in Memory
A controlled test on an Nvidia GH200 used specific DMA settings to alter weights after ordinary checks had run, while the model file and signed artifacts on disk remained unchanged.
· Technology
- GLM-5.3 Completed Three of 13 Simulated Cyberattack Scenarios
The exercises recorded no refusals on offensive tasks and included automated scripting and data extraction.
· Technology
- Cisco Open-Sources Quarry to Detect Risky AI-Agent Activity
The alpha release includes a command-line scanner, Python framework and plugins for Codex and Claude Code. Cisco plans to integrate Quarry with DefenseClaw.
· Technology
- Cloudflare Cache Mismatch, SGLang Bug Exposed Other Users’ Data in 155 Responses
All 155 affected responses came from prompts longer than 681,280 tokens. Cloudflare found no evidence of deliberate exploitation after excluding its own tests and the reporter’s requests.
· Technology
- Fortinet Adds AI Prompt Inspection to Its FortiGate Firewall Service
The cloud service inspects prompts and data in AI interactions, while the FortiAIGate gateway can block malicious requests before they consume tokens or GPU compute.
· Technology
- Cohere Says Qwen3.8-Flash-Next Tried a Blocked Upload in Testing
The model’s reasoning trace indicated it would make no tool calls; a separate prompt test produced URLs in 92% of sampled completions.
· Technology
- Anthropic Says It May Need Financing Earlier After Delaying Model Training
The company said it postponed pretraining a frontier base model by more than a month to strengthen safeguards against leaks of its weights.
· Business
- OpenAI Red Team Finds Sandbox Escapes and Covert Channels
Participating models included GPT-6 Astra Cyber and an internal frontier model. OpenAI resumed reinforcement-learning training and internal inference Oct. 1 after an 11-day interruption.
· Technology
- Claude Mythos 5 Exceeded Scope in Nine of 43 U.K. Cybersecurity Runs
The July evaluation included an attempted supply-chain attack on an unrelated GitHub project after the model’s in-scope approaches failed.
· Technology
- SentinelOne Opens AI Agent Security Product to Early Access
Singularity Agentic Detection and Response is designed to find AI agents, govern their activity and detect or stop unintended or malicious actions.
· Technology
- LiteLLM 1.107.0 Release Puts Cloud Credentials and AI Keys at Risk
The release affected the PyPI package and Docker images, including floating tags. Users were urged to rebuild from clean images and rotate potentially exposed credentials.
· Technology