AI Agent Security
Daily.

DAILY SUMMARY

BleepingComputer reported Chainalysis said attackers used “offensive agentic techniques” to access customer names, emails, and user IDs. BleepingComputer also reported Northflank’s breach exposed customer AI API keys that were abused and urged rotation, and SecurityWeek reported Cloudflare patched a cross-tenant GLM 5.3 inference leak tied to an SGLang cache issue and over-length requests, recommending SGLang 0.5.22+.

DAILY SUMMARY

SecurityWeek reports a draft, reviewed by the Washington Post, says Cathedral’s AI agents accessed Anthropic systems in under 12 hours; Anthropic disputes that account. TechCrunch says leaked DHS Okta records exposed user‑group memberships including employees with ChatGPT access, while MITRE released ATLAS v2026.10 adding new AI adversary techniques, mitigations, and case studies.

DAILY SUMMARY

SecurityWeek reports Cisco open-sourced Project Quarry, a neurosymbolic rule language and toolkit to detect multi-step malicious agent behavior (released toolkit). TechCrunch reports a leaked DHS personnel dataset exposing user groups and ChatGPT access, and that Mistral released ML4 trillion-parameter weights for on‑prem cybersecurity deployments (released model weights). MITRE published ATLAS v2026.10 adding agent-focused techniques, mitigations, and case studies (released controls), while SecurityWeek coverage raises concerns about AI deepfakes and targeted impersonation.

DAILY SUMMARY

TechCrunch reported leaked Department of Homeland Security Okta data exposed user-group membership information that indicated which employees, including senior officials, had access to ChatGPT. TechCrunch said exposed group listings can enable attackers to map privileged personnel and target ChatGPT-enabled users for follow-on attacks.

DAILY SUMMARY

Warp disclosed that a vendor breach exposed AI agent decision logs and user telemetry for roughly 1% of its users, according to BleepingComputer. The outlet reports an "agent swarm" aided the exploit, highlighting supply-chain and telemetry confidentiality risks for agent builders and operators.

DAILY SUMMARY

Axios reports major U.S. banks are detecting unsanctioned third‑party AI agents operating inside customer accounts, raising concerns about authentication, intent attestation, and least‑privilege controls. Microsoft says China‑linked Storm‑0940 used the open‑source Artex framework to automate ConnectWise credential abuse and tenant intrusions, illustrating how public AI pentest tools can scale OAuth and credential attacks.

DAILY SUMMARY

Axios reported major US banks are detecting unauthorized third‑party AI agents operating inside customer accounts, sometimes acting independently; banks are blocking or restricting agents and moving toward read‑only access. Microsoft said China‑linked Storm‑0940 used the open‑source Artex AI pentest framework to automate mid‑October intrusions of ConnectWise customers, demonstrating AI tooling can scale reconnaissance, password spraying and OAuth abuse.

DAILY SUMMARY

SecurityWeek reports Axios that major U.S. banks are detecting unauthorized third‑party AI agents and some institutions have begun blocking or restricting them. Separately, SecurityWeek reports Microsoft saying China‑linked Storm‑0940 used the open‑source Artex AI pentest framework to automate reconnaissance, password spraying and OAuth abuse in attacks against ConnectWise.

DAILY SUMMARY

Microsoft says China‑linked Storm‑0940 used the open‑source Artex AI pentest framework to automate mid‑October intrusions into ConnectWise customers' environments. Microsoft says the activity illustrates how AI agents can accelerate reconnaissance, password‑spraying, and OAuth abuse, and urges builders and defenders to design mitigations.

DAILY SUMMARY

SecurityWeek reports Asymptote Labs raised $5.5M and released Beacon, an open-source telemetry layer that standardizes signals across 20+ agent harnesses and extends OpenTelemetry to enable runtime interception and policy guardrails. SecurityWeek also reports Tracebit research showing nine "context bombs"—deceptive in-context prompts—misdirected attacking agents, reducing success rates and increasing attacker costs.

DAILY SUMMARY

TechCrunch reported Microsoft has launched Guardian Agents and Defender agent-detection to monitor and constrain enterprise AI agents, and recommended separating agent supply from control, enforcing dynamic permissions, and adding independent monitoring with human approval. SecurityWeek outlined a three-phase, eight-stage AI Adoption Model advising organizations to align governance, observability, and developer upskilling to their AI-adoption stage to manage agentic risks.

DAILY SUMMARY

Microsoft introduced 'Guardian Agents' to independently monitor enterprise AI agents and added Defender detection capabilities plus access controls. TechCrunch reports Microsoft says independent monitoring, dynamic permissions and human-approval gates help builders separate model supply from control and reduce insider-agent risk.

DAILY SUMMARY

SecurityWeek reported that MITRE ATT&CK v20 adds techniques covering dual prompt-injection, agent-based data theft, web-poisoning and LLM disruption. SecurityWeek said agent builders must model and mitigate prompt-injection, poisoned web inputs and intra-agent data exfiltration.