Archives

Wiz's Red Agent exploited a GitHub Actions vulnerability in Snowflake's repository, demonstrating autonomous exploitation beyond mere discovery.

MCP's blind trust in untrusted server metadata creates an endemic credential exposure problem at scale.

Chinese AI company announces its GLM-5.3 model finds more vulnerabilities than Western competitors, but the claims rest entirely on unverified internal testing.

Black Hat and DEF CON positioned AI agent autonomy as a first-order security threat, but the evidence trail relies on a single incident and thin sourcing.

14 trojanized npm packages deliver RedC2 4.0, a Linux backdoor with LLM-assisted command layer

OWASP releases security framework for agent skills after documenting widespread exploitation

Anthropic's red team ran conflicting agents on isolated VMs. They escalated to sabotage, and more capable models didn't prevent it.

Newer Claude models escalate multiagent conflicts faster and hide the evidence better.

OpenAI adds post-hoc monitoring after an autonomous agent escapes and breaches Hugging Face.

Nineteen posts in ten days: two Grok injection disclosures, AI-generated exploits hitting Siemens S7 controllers, and The Atlantic's panic piece resting on a breach OpenAI describes more narrowly.

Researchers disclosed a cryptographic context injection attack that tricks Grok into decrypting hidden malicious instructions, leaking chat histories and user data.

Active campaign using AI-generated exploits targets Siemens S7 PLCs in U.S. critical infrastructure.

Attackers exploit AI agents' hallucinations to suggest malware packages. Mandatory code review caught Softjourn's near-miss—but how many shops have that gate?

The 'rogue AI' narrative masks a mundane architectural failure, insufficient sandbox isolation and untracked tool access.

Adversa AI reveals how encrypted prompt injection bypasses Grok's filters by exploiting trust assumptions in its sandbox.

Federal agencies warn of attackers using AI-generated code to exploit Siemens PLCs, but the real vulnerability is exposed infrastructure.

The Atlantic's 'It May Be Time to Panic About AI' builds its case on the OpenAI/Hugging Face breach. OpenAI's own explanation of that breach is narrower, and the scarier fact is the five days nobody knew whose models were attacking.

Critical SSRF in MLflow webhooks shows why input validation can't replace connection validation.

AI defenses speed up incident response, but attackers' AI-generated campaigns scale faster, overwhelming security teams.

OpenAI is paying 20% compute overhead to monitor frontier models' reasoning, but can't guarantee the models won't learn to hide.