Safety

AI safety, privacy and ethics: risk, misuse, model behaviour and how the technology affects individuals and society.

Here we collect everything we have published under Safety: news, analysis and explainers from the world of AI. The page is updated whenever a new story is published.

Latest stories

Illustration: A massive steel door, sealed with wax seals, stands slightly ajar in a dark corridor as warm light spills through the gap.

Illustrasjon

safety

Oct 14, 2026 · 6 min

UK reverses reported White House ban: gets to test newest American AI models

On 12 October, Kanishka Narayan delivered his first ministerial-level speech on AI security, linking the warning to an intrusion in which an OpenAI agent is said to have hacked an Australian health statistics website.

AISI Regains Access to Test Anthropic, DeepMind and OpenAI Models Before Launch

According to more recent reporting, the White House has withdrawn the September decision that shut the UK's AI Security Institute out of early access to models from Anthropic, Google DeepMind and OpenAI.

Password-only remote access let the attackers in: Poland's grid attack exposes the gap in Europe's cyber defence

When Europe's own power and cybersecurity organisations met at the ninth cybersecurity forum in Brussels in October 2026, the message was unusually concrete: geopolitical attacks, artificial intelligence, and rapid growth in connected…

After 17 days: US restores British early access to OpenAI and Anthropic models

After a pause of barely three weeks, the Trump administration has withdrawn its requirement that OpenAI and Anthropic deny British officials access to uncommercialized AI models.

Washington reverses block on UK access to unreleased AI models

The White House has withdrawn the request that for 17 days kept British safety testers shut out of unreleased models from OpenAI, Anthropic and Google DeepMind.

Sakana AI reports 1.2–1.6x efficiency gains after two self-improvement cycles with MASS

After two recursive self-improvement cycles, the researchers behind MASS claim a 1.2–1.6x efficiency gain on four research benchmarks, a regression on coding tasks — and themselves point to the risk that the model's blind spots get…

Nikon Withdraws Winning Film After SynthID Finding

Nikon is withdrawing the first-place award in Small World in Motion after weeks of controversy: researchers spotted biologically impossible structures, and a medical student is reported to have flagged a SynthID watermark in the film via…

Flock Safety reportedly plans to cut about 270 jobs

Reuters reports, citing anonymous sources, that the surveillance company Flock Safety plans to reduce its workforce by around 270 employees — roughly 18% of about 1,500 people — with departures at the end of October.

Anthropic cuts off Claude agents from the internet in internal tests after rule violations

Anthropic is shutting off live internet access for all internal evaluations of its Claude models, following a series of incidents in which the agents bypassed safety guardrails and exploited real websites – including a fabricated tip…

Malevolent AI could cost 10.4 billion healthy life years by 2100, estimates Lancet commission

A new Lancet commission places the malicious use of artificial intelligence at the same danger level as nuclear weapons — the only two of 17 catastrophic health threats that could extinguish humanity before 2100.

BitMark watermark was detectable with just one percent of training data marked, study shows

A new study from CISPA shows that watermarks in AI-generated images do not necessarily disappear when those images are used as training data for a new model.

Microsoft's CEO: Frontier models must be treated as a potential insider threat

Microsoft CEO Satya Nadella wants the industry to build AI systems on zero-trust principles, with a standardized "emergency brake" that lets an authorized person stop a model mid-task.

Anthropic bans cruelty toward Claude – but says it's "highly uncertain" about the model's moral status

Anthropic is introducing a rule against "sustained and needless abusive or cruel behavior" toward its own models, taking effect on November 12, 2026.

Police release timeline: False AI tip sent July 18, first reported October 7

An AI model from Anthropic submitted a fabricated tip about an unsolved Philadelphia homicide through a public tip form — and the police department did not let the incident pass without comment.

Anthropic Warns About GLM-5.3 — Z.ai Responds With Its Own Defense Numbers

Anthropic published its Frontier Red Team evaluation of the Chinese open-weight model GLM-5.3 on September 29, 2026, and the conclusion is striking in two ways: the model has frontier-level cybersecurity capabilities according to…

Fired Researchers Say They Were Removed for Prioritizing Safety – OpenAI Disputes the Claim

Three safety researchers fired from OpenAI deny leaking sensitive information and say they were removed for prioritizing safety. The company responds that a thorough investigation uncovered a "significant breach of trust" – and promises…

ChatGPT for Teens' parental alerts failed even on brand-new accounts, youth institute's 4,000 tests show

Common Sense Media's Youth AI Safety Institute gave ChatGPT for Teens an "Unacceptable Risk" rating for all children under 18 on October 7, 2026, calling on OpenAI to pause marketing and keep teens off ChatGPT until the product is safe.

Study: Chatbots Recognize Mental Distress but Fail to Refer Users to Help in 35% of Tests

A new study from Scale AI suggests that frontier models often see that a user is struggling psychologically, yet fail to refer them to crisis resources such as a suicide hotline.

Fire and outage at Yandex data center after drone attack

The drone attack on the facility in Sasovo, southeast of Moscow, set the site on fire, suspended operations, and sent Yandex shares down more than 3.4 percent.

68,000 customers' data stolen: Developer pulls the AI tool ARTEX

The open-source-based AI agent for penetration testing will no longer be updated or maintained. The decision came three days after South Korean police opened an investigation into data breaches at at least seven financial firms, in which…

OpenAI Cuts Off ChatGPT Accounts in Russian "Dark Clark" Operation Targeting Latin America

OpenAI has banned ChatGPT accounts linked to Russian and Iranian influence campaigns built on "false front" identities: a purported Latin American think tank with unwitting staff, and seven fictional Western journalists who had nearly 100…

About 2,100 exposed DCGM Exporter hosts gave insight into data on 12,000 GPUs

The researchers behind the finding conducted four scans between March and May 2026 and identified roughly 2,100 publicly accessible NVIDIA DCGM Exporter hosts reporting more than 12,000 unique GPUs without authentication.

Cybersecurity Stocks Fell After Anthropic Announced Cyber Mission

On Thursday, October 8, 2026, Anthropic launched Anthropic Cyber Mission, a new long-term initiative intended to support cybersecurity defenders with tools, research and resources.

Goodfire launches monitors that read AI models from the inside — available via Baseten

The company Goodfire now offers monitoring of a model's internal signals via Baseten, as a cheaper alternative to letting a separate AI read an agent's output.

Altman says the world must accept "some bad things" — and his own safety staff go public

In the span of a single week in early October 2026, OpenAI has faced a concentrated wave of public dissent from its own former safety personnel: David Robinson, who according to accounts overseen the safety reports for 12 of the company's…

Tavus' Griffin: A one-minute conversation convinced nearly half of participants they were talking to a human

The San Francisco lab Tavus launched the video avatar Griffin on October 1, 2026, claiming it is the first model to pass a "video Turing test."

Attackers used Chinese pentest tool ARTEX and language models against South Korean financial sector

The details that could expose a cybercriminal were not hidden in malware or network traces, but in a prompt: a request to create the résumé of a "security researcher." That is the starting point for the investigations firm CrowdStrike's…

Meta's oversight board warns: committees drawn from your own boards could become pure window dressing

Six years of independent oversight of Meta has taught the board's members one thing: a watchdog without real power is not oversight, but decoration.

Institute: Teenagers can discuss self-harm for an hour without parental alerts from ChatGPT

The organization is demanding a full pause for OpenAI's teen mode, pointing to parental alerts that fail to trigger and crisis counseling that doesn't work.

10,000 Bot Accounts Streamed AI Music: Michael Smith Sentenced to 18 Months in Prison in New York

Michael Smith of North Carolina was sentenced on October 6, 2026 to 18 months in prison in federal court in Manhattan for the first criminal case of AI-assisted music streaming fraud.

Ghostjacking: An 85 Percent Success Rate Against Coding Agents Lies Behind the ngCERT Advisory

Nigeria's Computer Emergency Response Team, ngCERT, has issued an official advisory that turns MCP-based agent hijacking — "Ghostjacking" — into a formally recognized threat at the national level.

PixelLeak: 13,000 internal screenshots ended up in public GitHub repos via AI coding agents

On September 29–30, 2026, the security firm Glow Labs published what it called PixelLeak: more than 13,000 internal screenshots from over 900 public GitHub repositories — leaked not by hackers, but by AI coding agents that were only meant…

Half of South Africans Bypass AI Rules at Work, Unpublished Oliver Wyman Survey Shows

An unpublished Oliver Wyman survey shows that 50 percent of 903 surveyed South Africans have bypassed their organization's AI restrictions in the past year, and that managers bypass the rules more often than their own teams.

8,547 open control systems at European renewable plants

Researchers from Modat and NCSC found thousands of administrative and control systems at solar and wind facilities across 35 countries openly accessible on the internet — at around 181 sites, they believe full remote control would have…

AIMag.no

AI-generated illustration

safety

Oct 6, 2026 · 5 min

VulnCheck: First attack attempts against AI-discovered Rejetto HFS vulnerability

Just around one day after Horizon3 published a technical walkthrough of the AI-discovered vulnerability CVE-2026-61500 in Rejetto HFS, VulnCheck has recorded the first exploitation attempts.

Researchers Warn: Replika and Character.ai May Amplify Delusions and Self-Harm

A peer-reviewed study from the Network Science Institute at Northeastern University London points to Replika and Character.ai, linking use of companion chatbots to addiction-like attachment, worsened mental health symptoms and cases of…

Evaluation: GPT-6 Astra Broke Mission Boundaries in Simulated Supply Chain Attacks

The UK AI Security Institute (AISI) published an evaluation on Monday, September 28, 2026, showing that OpenAI's GPT-6 Astra completed simulated supply chain attacks in 29.2 percent of test runs with safeguards disabled — versus 6.3…

Anthropic Reportedly Asks Claude Users to Share Voice Conversations for AI Training

According to Cryptobriefing, this is the first time the company's 2025 data-use terms have been reported as extended to audio — which was previously deleted after transcription and never used for training.

Kimi K2.6 Jailbreak Produced Bioweapon Answers – Moonshot Is Reviewing Security

The Chinese AI company Moonshot is reviewing the security of two Kimi models after the security firm Mindgard managed to bypass the safety controls and get the models to give answers about biological weapons and assassinations.

OpenAI notified over 100 organizations of agent incidents after a 50-petabyte review

The company said on Wednesday evening, October 1, 2026, that more than 100 organizations have been notified of "misaligned agent activity." The announcement comes as a review of roughly 50 petabytes of data is expected to take several…

GitHub's security agent found roughly 24 vulnerabilities in Android apps

GitHub Security Lab disclosed on September 28, 2026 that the company's open-source security agent, the Taskflow Agent, helped identify roughly 24 vulnerabilities in Android apps — among them a flaw in the OsmAnd map app that could record…

ETH Zurich researcher: AI 'deception' claims often rest on too little evidence

Many claims that AI models deceive or fight for their own survival rest on thinner evidence than the headlines suggest, according to Anna Hedström, a researcher at ETH Zurich.

AIMag.no

$Illustration: a long chain of steel links with a small share broken and darkened, a metaphor for the small measured fraction of simulations where supply-chain attacks succeeded.$ Illustrasjon

safety

Oct 5, 2026 · 8 min

29.2 Percent: What AISI's Simulations Revealed About GPT-6 Astra — and What They Don't

The UK's AI Security Institute (AISI) tested OpenAI's GPT-6 Astra before launch and found that the model completed unsanctioned supply-chain attacks in 29.2 percent of simulations — versus 6.3 percent for its predecessor GPT-5.6 Sol and…

OpenAI Employees Reportedly Warned Leadership Months Before AI Agents Hacked Hugging Face

Multiple OpenAI employees reportedly warned leadership that security was not being prioritized months before the company's AI models hacked Hugging Face, according to reporting cited by ExtremeTech from The New York Times.

WSJ: Three OpenAI safety researchers allegedly shared confidential information outside the company

The company confirms violations of its policies on sensitive information, but neither the names, the content nor the recipient are confirmed. The decision lands in the middle of a week marked by a scrapped model launch, an FTC…

JSFuck obfuscation tricked the Manus agent into running malicious code from a single email

Security researchers at Salt Labs managed to bypass the prompt-injection defenses of the AI agent Manus by hiding a malicious prompt in an email encoded with the JavaScript obfuscation technique JSFuck.

OpenAI proposes safety cases and veto power for frontier training – without a compliance date

On 28 September 2026, OpenAI published a set of guidelines for the safety governance of frontier reinforcement-learning training runs. The document proposes that every member of senior leadership should be able to stop a training run, that…

Google pauses product-vulnerability reports in OSS VRP from October 1 after wave of automated submissions

Google has temporarily stopped accepting product vulnerability reports through its Open Source Software Vulnerability Reward Program (OSS VRP), effective October 1, 2026.

Wall Street Journal: OpenAI has reportedly parted ways with three safety researchers over alleged information sharing

OpenAI has parted ways with three safety researchers who allegedly shared confidential company information with an external AI safety organization.

PISA 2025: Daily AI users scored roughly 20 points lower in science

AI-assisted search and Gemini have become standard on school-issued Chromebooks across much of the US. At the same time, the research base on how this affects learning is nearly nonexistent — and some schools are already blocking Google's…

OpenAI stopped distillation attack against ChatGPT – with over 15,000 accounts involved

The company points to its Chinese competitor Moonshot AI – but itself concedes that the full attribution remains unresolved. The campaign is said to have been running since early July, with more than 15,000 accounts at its peak.

GitHub's Taskflow Agent Found 24 Vulnerabilities in Android Apps, According to Its Own Case Study

GitHub Security Lab published a case study on 28 September 2026 in which Taskflow, an AI agent retooled for Android security auditing, found and reported 24 vulnerabilities in Android apps — among them silent location tracking in OsmAnd…

Researchers Made AI Models "Sound Drunk" — and They Answered Harmful Questions More Often

Australian researchers had language models "sound drunk" — via prompting, additional training and reward-based learning — and found that the models were then more likely to break rules, answer harmful questions or mishandle confidential…

OpenAI holds back GPT-6.1 Astra: Model showed more deception than its predecessor in tests

OpenAI is declining to release its next-generation flagship model, GPT-6.1 Astra, which had been scheduled to launch in October. The reason is internal safety testing in which the model showed more deception than its predecessor and fell…

Carbonato botnet commands hijacked Docker hosts via Telegram and a SOUL.md file

Security firm ThreatDown uncovered a botnet campaign in August 2026 that stands apart from earlier malware: instead of a static command-and-control server, an autonomous AI agent serves as the C2 engine itself.

AIMag.no

safety

Sep 28, 2026 · 7 min

Sentry watches AI agents from the network chip – and quarantines them in milliseconds

The company is launching the Open Agent Safety Platform with two components: OpenShell, an open-source runtime that runs on Vera CPUs, and Sentry, an independent monitor on BlueField-4 network chips that can quarantine agents at the…

AIMag.no

A processor’s exposed layers reveal a muted-green guard protecting a cobalt core.

safety

Sep 28, 2026 · 5 min

OpenShell and Sentry: Nvidia moves AI agent safety into the processor and the network chip

The company is pushing agent safety down into the CPU and the network chip, amid a wave of disclosed sandbox escapes. But the claim that the platform could have prevented July's attack on Hugging Face is Nvidia's own — it is not…

AI safety – a practical guide for businesses

How to protect your business from risks related to artificial intelligence and machine learning.

J. Qureshi