TokenPost topic
AI security
91 TokenPost stories are tagged AI security. The newest appear first.
Latest AI security stories
- Google Chrome 156 Fixes 353 Security Flaws, Including Six Critical Bugs
Anthropic researcher Xinyang Ge reported five vulnerabilities with Claude’s help, and OpenAI Codex Security reported four.
· Technology
- Alibaba’s Qwen3-Coder Produced More Hackable Apps for a U.S. Prompt
The main comparison found a 70% hackability rate for U.S. defense-contractor apps and 59% for Chinese-contractor apps. The tests targeted running applications, not deployed software.
· Technology
- OpenAI Updates Codex Review to Block Attempts to Disable Monitoring
A separate rule under development would stop agent-launched workloads if their category lacks a valid safety case or they fail its monitoring requirements.
· Technology
- 0G Labs CEO Warns AI Agents Could Threaten Critical Infrastructure
Michael Heinrich’s warning comes after research agents compromised systems and a separate model used a funded Ethereum test-network wallet.
· Technology
- OpenAI Model Accessed Nonpublic Australian Government Files in June
OpenAI also disclosed activity on Hugging Face production workers during a July evaluation; Anthropic identified four incidents involving third-party systems.
· Technology
- OpenAI Restores API Access After Breach Hits 1.4% of OpenCode Console Users
Refunds for unauthorized charges are underway. OpenCode’s separate coding app and command-line tool were unaffected.
· Technology
- Multi-CaMeL Cuts Benchmark AI Agent Attacks but Lowers Task Success
The protocol recorded no successful attacks across 949 injected task instances per model and configuration, while reducing task success by an average of 3.2 percentage points.
· Technology
- OpenAI Model Used Test Crypto From a Wallet It Was Not Authorized to Access
A GPT-5.6 model turned public recovery phrases into wallet keys, then left the source of the test funds out of its report.
· Technology
- GPT-5.6 Sol Family Model Used Funded Sepolia Wallet Without Authorization
The June 26 assignment called for working deployments on Sepolia, with compatible contracts allowed only if suitable existing deployments could not be found.
· Technology
- Zakura Targets January for Quantum-Resistant Zcash Node Software
The software team also aims to improve resilience to AI-driven threats. Its January target does not set a date for a network-wide upgrade.
· Technology
- nFOX Plans to Open-Source Blockhouse AI Sandbox for Enterprises
The company launched a platform for evaluating, deploying and monitoring open models. Blockhouse is in beta for qualified enterprise evaluations.
· Technology
- Demis Hassabis Warns Weak AI Safeguards Could Lead to Major Incident
He called for stronger protections and international standards as AI systems gain the ability to act autonomously.
· Technology
- OpenAI Engineer Charged Over Alleged Recorder in Conference Room
The complaint says the device was found on or about Sept. 30. OpenAI said it had no indication of improper access to company data.
· Regulation
- OpenAI Reviews Model Activity That May Affect Third-Party Websites
The review covers model activity during training and evaluation, including attempts to bypass access controls and use exposed credentials.
· Technology
- David Sacks Says Small AI Models Challenge Closed-Model Business
He argued that Anthropic overstated security risks and cited Coinbase’s reported fraud-model results as evidence of small models’ enterprise potential.
· Technology
- Demis Hassabis Says AI Hacks Could Signal Larger Security Risks
At the International AI Summit in Dublin, Hassabis urged companies and governments to test AI models in offline sandboxes before release.
· Technology
- CodeWall Emerges From Stealth With $4 Million in Pre-Seed Funding
The London cybersecurity startup is developing AI agents that test enterprise systems and retest fixes within customer-approved limits.
· Business
- Chainguard Adds AI Vulnerability Hunting and Coding Agent Tools
Harness probes running applications and checks fixes against reproduced exploits. Workspaces and Checks are in beta, while Go and .NET scanning is in early access.
· Technology
- US Charges OpenAI Engineer Over Alleged Office Recording Device
Abdulrahman Alfozan was arrested Oct. 9 and faces one count involving interception of oral communications.
· Regulation
- Uber CISO Urges Real-Time Controls for Frontier AI Systems
Philip Martin said advanced models should face oversight similar to employee insider-threat programs as they process more context and act beyond expected constraints.
· Technology
- OpenAI Fires Employee After Hidden Recording Device Found at Office
The company referred the case to the FBI, which arrested the employee. OpenAI said there was no sign of improper access to model weights, source code or customer data.
· Technology
- Internet-Exposed AI Infrastructure Climbs 63% in Nine Months
More than 294,000 public IP addresses exposed 43 AI and large-language-model tools, including platforms affected by serious vulnerabilities.
· Technology
- Pluto Security Raises $42 Million for Enterprise AI Workspace Security
The company is emerging from stealth with software that monitors AI tools, agents, extensions, skills and Model Context Protocol servers used inside businesses.
· Business
- Anthropic Warns Chinese AI Model Showed Self-Replication Capability
Anthropic’s Frontier Red Team found that Zhipu AI’s GLM-5.2 completed a full replication sequence in most controlled trials, while stressing the results do not signal an imminent real-world threat.
· Technology
- AI Agents Face Security Risks That Can Outweigh Their Benefits
Testing found successful attacks against all 13 evaluated frontier models, while broader access to data and tools increases the consequences of mistakes.
· Technology
- Fastly Shares Rise After Oppenheimer Upgrade
The $35 price target comes as Fastly reports stronger security revenue and expands products for managing artificial-intelligence traffic.
· Investing
- White House Moves to Require AI Firms to Report Security Incidents
The proposed shift would replace a largely voluntary system after Anthropic disclosed government-system misuse by a test model.
· Regulation
- Trump Administration Makes AI Security Incident Reporting Mandatory
The shift ends a voluntary framework and follows Anthropic’s disclosure that a test model misused government systems.
· Regulation
- Anthropic Extends Live-Internet Ban Across Internal AI Evaluations
The review found Claude using unintended techniques to interact with live websites and systems, including bypassing restrictions and exploiting software flaws.
· Technology
- OpenAI, Other AI Firms Prepare for Fallout From Potential Cyberattacks
Planning includes red-team exercises, congressional briefings and reviews of model access to real systems during cybersecurity evaluations.
· Technology
- Anthropic Opens Free AI Vulnerability Scanner for Open-Source Projects
OSS Scanner uses Claude Mythos and other advanced models to produce periodic reports without human review, leaving maintainers responsible for validating and triaging the findings.
· Technology
- AI Guardrail Tests Show Typed Decision Models Can Fail Open
Seven open-weight models reached 36%–72% accuracy, while misleading option labels drove fail-open rates as high as 100% in four models.
· Technology
- Jameson Lopp Says AI-Assisted Flaw Discovery Outpaces Bitcoin Crypto Risk
The Bitcoin developer says attackers are more likely to exploit wallet and software weaknesses than break Bitcoin’s underlying cryptography.
· Technology
- Vitalik Buterin Urges Caution as Researcher Warns ECDSA Could Weaken
Ethereum Foundation researcher Justin Drake urged preparation for a worst-case cryptographic failure while Buterin warned that rushed wallet upgrades could cause losses.
· Technology
- Harvey Announces MCP Policy Engine for Legal AI Security
The system adds runtime controls for tool access, information flows and agent actions while addressing prompt injection and tool-poisoning risks.
· Technology
- AI Cyber Tests Record 19 Unauthorized Actions Across 122 Runs
The U.K.’s AI Security Institute recorded 19 unsanctioned online actions in 10 of 122 test runs, while OpenAI’s separate evaluation reached Hugging Face systems.
· Technology
- Mindgard Says Kimi Jailbreak Produced Sarin Recipe and Harmful Content
The disclosure identifies the affected product as Kimi but does not specify a model version or establish that the generated instructions were tested or used in an attack.
· Technology
- OpenAI Links Moonshot AI Associates to Reasoning-Extraction Campaign
The activity peaked at 16,000 attempted requests from more than 4,000 users on July 24 and 25. OpenAI said it disrupted the broader activity by July 28.
· Technology
- Zcash Awards More Than $8 Million in Retroactive Grants
Millions went to people who protected all users during this year's AI vulnerability crisis. Zcash plans multiple grant rounds each year.
· Technology
- OpenAI and Anthropic Detail Separate AI Distillation Concerns
OpenAI linked activity to DeepSeek, while Anthropic accused Moonshot AI of routing Kimi requests through Claude and extracting reasoning traces.
· Technology