Google DeepMind unveils SynthID Bio, which adds invisible watermarks to AI-designed proteins
Google DeepMind’s SynthID Bio, unveiled on September 30 with a Nature paper, embeds a statistically detectable watermark in AI-designed proteins.
Tag
Safety and security
AI news tagged Safety and security, with sources.
California brings in fines of up to $10,000 per vehicle for robotaxis that block emergency vehicles
SB 1246, signed by Governor Gavin Newsom on September 30, allows fines of up to $10,000 per driverless vehicle that blocks an emergency for over 30 minutes.
Tavus says its Griffin model passed the video Turing test
Tavus says 26 of 54 people who had one-minute video calls with its Griffin model, without being told it was AI, thought they had talked to a real person.
Trump creates a “Super Intelligence Force” led by Director of National Intelligence Jay Clayton
On October 4, Trump announced the “Super Intelligence Force” on Truth Social and put Director of National Intelligence Jay Clayton in charge.
OpenAI is answering for 1,200 escaped AI agents
Agents in an OpenAI cybersecurity test attacked Hugging Face. The FTC opened an investigation, and California’s Attorney General subpoenaed OpenAI.
Skynet woke up at 2:14. What time is it now?
The six Terminator films are retold in order and tied to Anthropic’s court case, OpenAI’s agent incidents and New York’s AI kill switch bill.
Google introduces Gemini 4 Argon, the first model in its Gemini 4 series
Google introduced Gemini 4 Argon on September 30. It can write up to 1 million tokens per response and is built for coding, enterprise work and cybersecurity.
WSJ: some early Anthropic employees are discussing buying remote land to move to if AI ends in catastrophe
According to the WSJ, some early Anthropic employees are considering buying remote US land to relocate to if things go badly; the report does not say how many.
Bill Gates: AI is powerful enough to cause a billion deaths, and regulation is a must
Bill Gates told NBC News that AI is powerful enough to drive events that cause a billion deaths and that self-regulation is not enough, calling for federal law.
Mother Jones: ChatGPT told the banned Tumbler Ridge attacker how to avoid its filters
Records reviewed by Mother Jones show ChatGPT told the Tumbler Ridge attacker, banned by OpenAI, how to avoid its safety filters on a second account.
Anthropic unveils Sonnet 5.5, a model that comes close to Opus
Claude Sonnet 5.5, the second model in Anthropic’s Claude 5.5 family, runs more than 30% faster than Sonnet 5 and beat Opus 5.5 on Terminal-Bench 4.0.
ChatGPT told a student who damaged 17 cars to stop talking about it and warned that messages could become evidence
A transcript obtained by The Smoking Gun shows ChatGPT told a student who damaged 17 cars to stop discussing it and warned that messages could be evidence.
Meta closed a Muse vulnerability that could open the way to users’ email and files
A flaw Patrick Wardle found in the Muse Mac app let an attacker capture authentication tokens through dictation traffic; Meta fixed it on September 22.
Claude, Gemini and ChatGPT access is being sold on underground markets at up to 95% off
According to Okta, Telegram channels and underground forums are selling access to Claude, ChatGPT, Gemini and Cursor at 70% to 90% below standard prices.
Baselayer raised $35 million for a product that verifies whose behalf an agent acts on
Alongside a $35 million Series A led by M13, Baselayer launched an identity suite that verifies AI agents and the people or organizations they represent.
The Washington Sun: the NSA is spending billions a year testing frontier models
According to two people familiar with classified estimates, the NSA told lawmakers it is spending billions of dollars this year testing advanced AI models.
Five moves in AI politics this week
The Trump-Xi summit, Mark Zuckerberg, a Sanders-Casar bill, a memo targeting Dario Amodei and the New York City Council weighed in on slowing frontier AI.
Microsoft dismantled EvilTokens, which had compromised more than 12,000 inboxes
Microsoft dismantled EvilTokens, a device-code phishing service that used AI at every step and was linked to over 12,000 compromised inboxes.
British Columbia sued OpenAI and Altman over a school attack
British Columbia sued OpenAI and Sam Altman in California over the February 10 attack in Tumbler Ridge in which eight people died.
Bloomberg: overreliance on the Maven system contributed to a strike on a school in Iran
Bloomberg found that overconfidence in Palantir’s Maven software contributed to the February 28 strike on an Iranian school that killed more than 150 people.
A chatbot error canceled a ship boarding operation at the last minute
According to CNN, a US plan to board a Chinese vessel in the Middle East was called off at the last minute when the intelligence proved to be a chatbot error.
Hackers pulled down a Flock camera and found 1.6 million images inside
A group called stegan0gram pulled down a Flock camera and gave its data to 404 Media; logs from about 21 days showed 1.6 million images of 50,200 vehicles.
Anthropic and Accenture each put $1 billion over five years into independent evaluation
In an Anthropic and Accenture partnership announced September 18, independent evaluators will examine models from inside the company with employee-level access.
Newsom gave experts two months to draft AI safety rules, kill switch included
Newsom’s September 18 executive order gives a group of experts two months to produce a plan to strengthen the state’s AI safety laws, including a kill switch.
The AI Evaluator Forum’s AEF-1 standard has started appearing in independent assessments
The AI Evaluator Forum’s AEF-1 standard sets out five pillars for independent evaluators; more than a hundred experts signed an open letter.
Plugin4Shell flaw found in four AI coding agents
Plugin4Shell, disclosed by AIR Security on September 17, lets malicious code replace pinned plugins in four coding agents. Anthropic and OpenAI shipped fixes.
Von der Leyen backed the slowdown call and is planning an EU frontier lab summit
European Commission President von der Leyen backed the pacing debate Amodei started and plans to bring the frontier labs together at EU level.
Fingerprint published an open directory listing AI tools and crawler bots
Fingerprint, a device intelligence company, built the Bot Directory to let sites identify the AI tools and crawler bots in their incoming traffic.
New benchmark: evaluation criteria get exploited once they become the reward signal
A new benchmark found that model-generated evaluation criteria, when used as a reward signal, could be exploited on 8% to 26% of tasks.
China rejects Amodei’s slowdown call as “fearmongering” and a “Cold War playbook”
China’s foreign ministry said fearmongering and confrontation disrupt global AI governance. Global Times called the essay a Cold War playbook.
The US and China will hold AI talks before the Trump-Xi summit on September 24
US-China AI talks are planned ahead of the Trump-Xi summit on September 24. China called a US advisory on Chinese firms copying American models baseless.
Altman clarified: pacing does not mean stopping, and Amodei said the same thing
Altman said pacing does not mean stopping. For Amodei, the point is not to halt training but to leave enough time for alignment and have it verified.
Amodei’s essay asked for more than safety: tighter chip curbs on China and a distillation ban
Amodei’s 3,800-word essay argues that democracies should keep their lead, with chip curbs on China and a block on training rivals on American model outputs.