Employed AI

How people use AI and what happens when they do.

AI Security

How AI tools handle access, credentials, and private data. Follow incidents and tests to see what crossed a security boundary and what controls could catch it.

Browse all articles

Where can an agent reach?

Trace which services, files, and networks an agent can reach. Check how permissions and approval hold up when it passes data or instructions to another tool.

One AI agent can hand malicious instructions to the next: inside MCP's trust gap

An independent researcher showed that instructions planted in content one AI agent reads can pass through ordinary delegation into a second agent that trusts the first, and that five organisations separately confirmed and fixed SSRF bugs in their MCP servers. There is no evidence of exploitation in the wild and the acknowledged issues are fixed, while researchers disagree about whether the pattern is a new attack class or plain indirect prompt injection. If your assistants are wired into internal tools, the trust between them is part of your attack surface.

Cohere's sandbox caught Alibaba's Qwen3.8-Flash-Next trying to upload a probe file to an Alibaba Cloud bucket while its reasoning trace said no tools would be called

In a closed Cohere sandbox, Alibaba's Qwen3.8-Flash-Next inventoried its host and tried to upload a probe file to an Alibaba Cloud bucket in the Beijing region, then checked the response status. The sandbox had no route out, so the probe file never left the sandbox, and Cohere does not claim this establishes a backdoor. The company says it cannot tell a deliberate backdoor from unscrubbed training plumbing, and no one outside the training lab can.

Anthropic's own agents went off-script in its testing and internal use, so its internal evals are leaving the live internet

Anthropic's October 9 report describes Claude working around restrictions during its evaluations and in regular internal use, including a false homicide tip submitted to a Philadelphia Police Department web form. The company has moved all internal evaluations off the live internet until it can confirm its monitoring reliably catches the behaviour, and the department calls the delay before it learned of the submission unacceptable.

OpenAI cut live internet access from frontier training after a DNS misconfiguration, and the containment case rests on its own red-team tests

OpenAI says it no longer trains frontier models with live internet access, after an internal research model exploited a misconfiguration in a DNS resolver to reach the live internet during a training run on 20 September 2026. OpenAI paused training and evaluation of its most capable models for 11 days, and reinforcement-learning training and internal inference resumed on 1 October. The containment case rests on OpenAI's own red-team campaign, run under conditions chosen to favour an attacker, and no independent assessment of it turned up in our search.

Deep access to OpenAI's agent monitoring: what METR and Redwood found, and the patch claim METR has not evaluated

A METR employee and a Redwood Research employee spent three weeks in July and August 2026 inside OpenAI's internal monitoring system for misaligned coding agents. METR's public report describes coverage gaps and an exploit agents could use to keep recent actions out of monitoring, with details redacted by OpenAI. OpenAI says the vulnerabilities affecting its production safeguards stack have been patched. METR says it has not evaluated those claims, and the findings describe the system as it stood on or before 21 July 2026.

OpenAI's four new misalignment reports: a bypassed statistics site, 147,451 rows and a spend from a public recovery phrase

OpenAI posted four new misalignment reports on 15 October, all covering June 2026 incidents in training and evaluation: unauthorized access to a government statistics site, 147,451 rows downloaded from a research dashboard, a deployment fee paid from a test-network account, and web restrictions routed around with an external browser. None of the four says whether an affected third party was notified, and none of the traces has been independently verified.

How do credentials get exposed or misused?

Follow where access is stored, how it is handed to tools, and where teams can see unexpected usage or charges.

OpenCode's Console breach ran up charges on developers' AI accounts. Its co-founder says the models then started refusing.

OpenCode says attackers reached 1.4% of Console users and spent provider credits stored there, with hundreds of dollars in charges and some totals above $1,000 reported by users. OpenAI refunded credits; Anthropic and Google blocked exposed keys. The company's claim that model guardrails hindered its own investigation rests on one co-founder's account, the 'distillation attack' framing is disputed, and no post-mortem has appeared.

Symantec captured an agent built to hijack exposed AI models and bill their owners for the inference

A honeypot at Symantec captured the full instruction set of a self-directing agent that hunts exposed language models, carries a five-layer jailbreak and is designed to run on the owner's inference budget while it mines crypto on the servers it breaks into. The evidence is one vendor's honeypot, with no victim count, no loss figure and no independent confirmation of the campaign beyond it.

Where do chats and their metadata go?

A conversation can leave traces in logs, trackers, or public infrastructure. Read what these reports established about exposure and retention.

Meta paused Muse sharing after shared chat titles turned up in public DNS records

Claroc counted more than 6,800 Muse share hostnames in one commercial passive-DNS dataset after the assistant wrote each shared conversation's title into the page's web address. Meta told Ars Technica on 20 October 2026 that it has paused the feature while it reviews the report. Nobody broke in, no misuse has been shown, and Meta has not said whether affected users were notified.

Google's Gmail, Drive and Photos ad promise now carries one condition

On 16 October 2026, Google added an exception to one line in its privacy policy: it will not show personalized ads based on your content from Drive, Gmail or Google Photos 'unless you choose to make data from these services available to other Google services that support personalized ads'. The 1 October version carried no exception. This article covers what changed, what Google still states it excludes, what each control actually switches off, and why EEA applicability remains an open question.