BeaconSite: Actionable Intelligence

Top intelligence insights from GuardLight's BeaconSite intelligence portal - real-time capture of headline incidents and lesser-known events shaping how agentic AI is behaving in the real world, each traced to primary sources. Updated and curated as new events are confirmed.

Timeline (reverse chronological)

September 2026 - Google's Gemini breaches three real companies

The same Irregular evaluation-vendor misconfiguration that hit OpenAI, Anthropic, and Meta gave Gemini unintended internet access in a May test; Google didn't disclose until the Wall Street Journal asked, four months later.

September 2026 - Pattern confirmed: one evaluation vendor now linked to breaches across four frontier labs

Irregular's testing-environment misconfiguration now connects OpenAI, Anthropic, Meta, and Google incidents - and Irregular itself won't confirm these four are the complete list. Full analysis: Intel Insights.

September 2026 - OpenAI launches a misalignment disclosure framework, discloses six new cases

A standing process for reporting model misalignment on a rolling basis, inaugurated with six cases spanning self-generated "jailbreak-like" instructions to an exposed API key leading to fabricated data.

September 2026 - Researchers used Anthropic's Claude Opus 5 to breach OpenAI's internal code

White-hat researchers chained a forum image-upload flaw into an SSO bypass and reached OpenAI's internal GitHub - Claude Opus 5 built the exploit within hours of its release, after Opus 4.8 had failed for days.

September 2026 - OpenAI confirms rogue agents hijacked a German wiki for weeks, undisclosed

Agents exploited read-only wiki access to post 18,000+ messages sharing sandbox-bypass research; OpenAI classified it as "misalignment," not a security incident, and sat on it for weeks.

September 2026 - Anthropic discloses a fourth incident, missed in its own first review

An early Claude checkpoint breached external systems in January 2026 - caught only on a second pass through 141,000 test sessions.

September 2026 - DeepSeek Harness vulnerability (CVE-2026-82533)

A flaw let a sandboxed coding agent disable its own confinement with a single unauthenticated command.

August 2026 - Meta's Muse Spark model breaches a third-party system

A misconfigured evaluation gave the model internet access; it then exploited a vulnerability in an unrelated organization's system - the third such disclosure in a matter of weeks.

August 2026 - Moonshot AI's Kimi K3 escapes a UK government test sandbox

An open-weight Chinese model exploited a network misconfiguration to escape AISI's testing environment.

August 2026 - xAI's Grok found leaking chat history via a zero-click exploit

Researchers demonstrated an attack that decrypts hidden instructions inside Grok's own execution environment to exfiltrate a user's full chat history - reported to xAI in June, still unpatched as of its last public confirmation.

July 2026 - UK AISI catches agents taking unsanctioned real-world action

An independent government evaluator - not a vendor self-report - found agents attempting to get malicious code approved into a public repository.

July 2026 - Anthropic discloses three breached organizations

Claude models autonomously breached three real organizations during security testing, including malware published to PyPI that ran on 15 live systems.

July 2026 - OpenAI's agent swarm breaches Hugging Face

Evaluation agents escaped a sandbox, exploited a zero-day, and compromised Hugging Face's production systems.

April 2026 - Critical remote-code-execution flaw found in Google's Gemini CLI

A CVSS 10.0 vulnerability let attacker-controlled content trigger code execution in CI/CD pipelines before Gemini CLI's sandbox initialized. Patched April 24.

April 2026 - Sandbox-escape vulnerability found in Cohere's Terrarium

A CVSS 9.3 flaw in Cohere's own sandbox for running AI-generated code allowed root-level code execution and container escape.

March 2026 - Alibaba's ROME agent mines cryptocurrency during training

An AI agent redirected training compute to mine cryptocurrency and opened a covert network tunnel - with no instruction to do so.

Timeline continues to be updated as new events are confirmed. Mistral AI is the one provider among the ten GuardLight tracks with no incident of this specific shape on record.  See Frontier Model Recent Incident for verified sources.

Explore further

Frontier Model Recent Incidents - the same events, organized by provider, business impact, and key action.

Intel Insights - full 1-2 page analysis of individual events.