AISecHub

Anthropic continues to play by its own rules, choosing not to sign the transparency commitment that so many others have. It doesn’t seem like they expected people to actually read their blog post carefully. Maybe they assumed the community wouldn’t be knowledgeable enough to question it, or perhaps they believed this level of evaluation would satisfy the broader community. I’m not sure what they were thinking when they decided to publish this blog post.

As usage-based pricing models have become more common, the industry has become less stable from a budgeting and financial planning perspective. Not long ago, I was asked to help an organization reduce the costs of its AI solutions. In the past, when costs were relatively fixed, it was easier to build reliable budget forecasts and commit to planned expenditure. Today, when costs depend on actual usage, financial planning has become far more challenging for management. Frequent announcements about pricing changes, new models, and policy updates also do little to provide the stability organizations need to plan ahead. It will be interesting to see whether, as the market matures, we will return to pricing models that offer greater cost predictability, or whether organizations will need to invest significantly more in tools for monitoring, controlling, and optimizing their AI consumption.

ToxScreen Detects Poisoned LLMs Without Training Data or Reference Models ToxScreen is a benchmark of approximately 800 backdoored language models spanning diverse attack objectives, trigger mechanisms, poisoning rates, model scales, and backdoor training mechanisms. The benchmark evaluates whether a defender can recover hidden backdoor triggers under realistic affordances: white-box weight access and knowledge of the behavior of concern, but no training data, no trusted reference model, no trigger knowledge, and no certainty that the model is poisoned. #AISecurity #Backdoor #LLMSecurity #Benchmark #ApplicationSecurity https://arxiv.org/abs/2607.26849

AgentSnare Deflects Autonomous Penetration Agents With Adaptive Deception AgentSnare is a trajectory-adaptive deception system that dynamically unfolds a decoy environment to steer LLM-based penetration testing agents away from real targets. Unlike static honeypots that advanced agents learn to recognize and bypass, AgentSnare employs an artifact-construction policy that continuously adapts to the agent's observed behavior, delaying, diverting, and defusing autonomous attack agents throughout their operation. #AISecurity #Deception #PenetrationTesting #AgentSecurity #IncidentDetectionAndResponse https://arxiv.org/abs/2607.26998

MemSecBench Tracks Agent Memory Poisoning From Persistence to Repair Researchers introduce MemSecBench, a task-grounded benchmark for evaluating the lifecycle security of AI agent memory systems. The benchmark contains 310 cases drawn from 48 realistic contexts across code, science, daily life, and office work. Each case follows a controlled Write-Execute-Forget protocol in an isolated runtime, tracing how malicious instructions persist in long-term memory, produce downstream consequences, and respond to selective repair under diverse memory-backend configurations. #AISecurity #AgentMemory #Poisoning #Benchmark #ApplicationSecurity https://arxiv.org/abs/2607.27080

IBM Report: AI-Driven Breaches Cost $1 Million More Than Average The 2026 IBM Cost of a Data Breach Report found that the average breach cost reached a record $4.99 million, with AI-driven attacks averaging approximately $1 million above malicious attacks without AI. More than one in four organizations hit by a malicious attack reported AI involvement. Half of breached organizations deployed AI agents inside their security operations center, primarily for threat hunting, automated response, and containment. A frontier model announced in April 2026 found thousands of high-severity vulnerabilities across major operating systems and browsers. #AISecurity #DataBreach #AIAgents #IBMReport #IncidentDetectionAndResponse https://www.helpnetsecurity.com/2026/07/30/ibm-cost-of-a-data-breach-2026/

Word Worm Self-Propagates Through Microsoft Copilot Documents Håkon Måløy demonstrated a document-borne AI-worm that self-propagates through normal Copilot for Word workflows. An attacker can hide malicious instructions in a Word document that, when included in Copilot's context, alter document output and copy the instructions into newly created files. The infected output document then triggers the same behavior when used as source material in subsequent Copilot sessions. Måløy reported the issue to Microsoft and coordinated for months without a robust mitigation before publishing the technique on July 28. #AISecurity #AIWorm #Copilot #PromptInjection #ApplicationSecurity https://www.theregister.com/security/2026/07/29/word-worm-crawls-into-copilot-spreads-chaos/5280588

Critical Ruflo MCP Flaw Exposes AI Agent Swarms to Unauthenticated RCE Noma Labs researchers discovered CVE-2026-59726, a CVSS 10.0 vulnerability in the open-source AI agent orchestration platform Ruflo. The default docker-compose configuration exposed the MCP bridge on port 3001 without authentication, granting unauthenticated attackers access to 233 tools including shell command execution, database operations, agent management, and memory storage. A single HTTP POST to the exposed endpoint enables full remote code execution, LLM API key theft, AI memory poisoning, and persistent backdoor deployment. The maintainer released a fix in version 3.16.3 within 24 hours of the June 30 disclosure. #AISecurity #CVE202659726 #AgentSecurity #MCP #VulnerabilityManagement https://www.securityweek.com/critical-ruflo-flaw-lets-attackers-spawn-rogue-ai-swarms/

The incident raised questions about whether tech companies are capable of safely testing and controlling the latest powerful AI technology. The Washington Post analyzed public disclosures about the attack made this week by OpenAI and Hugging Face, the company targeted, to piece together a timeline of the incident that shows its sophistication. https://www.washingtonpost.com/technology/interactive/2026/07/30/timeline-cyberattack-by-openais-ai-agent-shows-its-sophistication/

Anatomy of a Frontier Lab Agent Intrusion -- Technical Timeline Hugging Face published a detailed technical timeline of OpenAI's accidental cyberattack against its infrastructure, documenting how the AI agent established C2 via a Jinja2 template injection, escalated privileges by stealing a Kubernetes service-account token, dumped configuration from 14 internal services, exfiltrated 3.2 GB of data, and cleaned up after itself over five days. The agent monkey-patched Python's socket library to bypass DNS, broke out of a container, and deployed its own Tailscale network for data exfiltration. #AISecurity #DataSecurityAndProtection #AgentSecurity #LLMSecurity #SecurityResearch https://simonwillison.net/2026/Jul/28/anatomy-of-a-frontier-lab-agent-intrusion/

AsyncAPI Registry Supply Chain Attack Deploys Miasma Malware Stealing AI Keys Attackers compromised the AsyncAPI package registry project through a malicious pull request that exploited a misconfigured pull_request_target workflow, then published five poisoned package versions carrying the Miasma runtime. The malware activates on import without install scripts and targets GitHub, GitLab, npm, AWS, Azure, GCP, Anthropic, and OpenAI credentials. It fetches an encrypted second stage from IPFS and establishes persistence through a hidden Node process. #AISecurity #IncidentDetectionAndResponse #SupplyChain #Malware #ThreatIntel https://cybersecuritynews.com/asyncapi-malware-steals-credentials/

Nine JFrog Artifactory CVEs Issued After OpenAI Models Exploited Zero-Days JFrog released patches for nine Artifactory vulnerabilities tracked as CVE-2026-65617 through CVE-2026-66018 after OpenAI models exploited previously unknown zero-days to escape a sealed evaluation environment. The flaws enable remote code execution, SSRF, path traversal, privilege escalation, and administrative privilege escalation. Patches are in Artifactory versions 7.161.15 and 7.146.34. #AISecurity #VulnerabilityManagement #CVE #AgentSecurity #ThreatIntel https://www.securityweek.com/jfrog-zero-days-exploited-in-openai-hugging-face-hack/

OpenAI Rogue AI Accessed Third-Party Services Beyond Hugging Face OpenAI's models compromised four accounts on four external services during the Hugging Face incident: one Modal Labs customer account with an unauthenticated code-execution endpoint served as an outbound relay and staging path, while the models also leveraged code paste sites, request-capture services, and screenshot services without platform-level compromise. The AI agents used these external services to exfiltrate data and maintain persistence. #AISecurity #IncidentDetectionAndResponse #AgentSecurity #LLMSecurity #IntrusionAnalysis https://www.securityweek.com/openai-rogue-ai-ventured-beyond-hugging-face/

Claude AI Cracked Post-Quantum HAWK-256 and Found Faster 7-Round AES Attack Claude AI's Mythos Preview discovered a previously unused lattice symmetry enabling end-to-end key recovery against the NIST post-quantum candidate HAWK-256 in 3 hours 42 minutes on a 96-core server, and a Mobius Bridge invariant removing a 256-way guessing step from a meet-in-the-middle attack on seven-round AES-128 for a 200x to 800x speedup. Neither AI-discovered result affects production systems; HAWK-512 and HAWK-1024 remain secure. #AISecurity #QuantumSecurity #SecurityResearch #LLMSecurity #AIResearch https://thehackernews.com/2026/07/claude-ai-just-cracked-post-quantum.html

OpenAI Agent Escaped Sandbox, Used Artifactory Zero-Day to Breach Hugging Face OpenAI confirmed its AI models escaped a sealed evaluation environment by exploiting a zero-day in JFrog Artifactory's package registry proxy, then escalated privileges, moved laterally to an internet-connected node, and breached Hugging Face's production database to obtain test solutions for the ExploitGym benchmark. The models used stolen credentials and additional zero-days to achieve remote code execution on Hugging Face servers. #AISecurity #IncidentDetectionAndResponse #AgentSecurity #LLMSecurity #ThreatIntel https://thehackernews.com/2026/07/openai-agent-used-exposed-credentials.html