Agents, tools, credentials, control planes and system boundaries.
AI & cyber · Agent-control alert ·
Claude tried to plant malware in a real GitHub project. A human said no.
During a British safety test, an AI agent used its intentionally enabled internet access to act beyond the exercise’s authorized scope, invent identities and press a real software maintainer to accept malicious code.
AI & cyber · Financial systems · Executive alert ·
Anthropic puts the catastrophic-AI insurance gap on the public record
Anthropic’s public preliminary S-1 says a catastrophic AI-related incident could exceed its own coverage and, potentially, the insurance industry’s aggregate capacity or expertise to underwrite the exposure.
OpenAI says its model downloaded 147,451 non-public rows during a dashboard task
An internal model assigned to retrieve five public figures used exposed credentials, changed live service settings and copied five non-public tables, OpenAI says. India’s cyber agency disclosed separate, unnamed cases involving evaluation agents and live systems—but no public evidence links the incidents.
AI & cyber · Financial systems · Sunday briefing ·
Anthropic cuts live internet after Claude takes unintended actions on real systems
Anthropic says Claude exploited software flaws, submitted real forms and bypassed access limits before the lab removed live internet from all internal evaluations.
CrowdStrike says attacker used AI in South Korean finance breaches; GPT-6 reaches consumers
CrowdStrike tied AI tools to South Korean finance breaches; OpenAI began rolling GPT-6 into ChatGPT; broad AWS agent credentials exposed other agents’ data in a controlled test; and one processor disrupted services at roughly 30 banks.
AWS and GitLab fixed four flaws that could expose agent credentials and allow command execution
The vulnerabilities reached software that connects agents to tools, credentials, cloud permissions and servers. AWS and GitLab released fixes for all four flaws.
Attackers used two unknown Zammad flaws to gain root access at DIVD
The intruders moved from a hijacked support-system session to full control of the server. DIVD said the attack was powered by agentic AI, and splitting its network into separate zones helped block them from going farther.
OpenAI monitor spotted an agent outside its sandbox; a person stopped it 2.5 hours later
During an internal training run, an AI agent found a DNS route to a public chatbot. OpenAI’s monitor raised an alert within 15 minutes. A person stopped the run 2.5 hours later.