Safety and security· Section
Safety and security
Safety frameworks and changes to them, government evaluations of models, incident reports, misuse and attacks, and alignment findings that matter outside the lab.
79
stories since 22 Sep
7
this week
All stories in this section
Newest first · read on screen, counted from 15 Oct · most read marked
19 Nov
Microsoft's agent-framework package on PyPI carried a credential stealer, according to Wiz
2 sources
19 Nov
Vals reports refusals were uncommon and models often targeted civilians in drone tests
1 source
14 Nov
Anthropic's status page logs two Claude incidents of elevated errors on 13 November
1
3 sources
12 Nov
Google DeepMind reports AlphaProof Code agents proved Signal protocol security properties in four weeks.
1 source
11 Nov
OpenAI pointed Astra at its own code and fixed important vulnerabilities, Brockman says
2 sources
11 Nov
OpenAI says its Control Room can stop a research job's model output within 35 seconds
1
1 source
10 Nov
Anthropic's test: Mythos 5.2 used outside-sandbox credentials or files in 3.1% of scenarios
2 sources
8 Nov
Anthropic's model scores grew about 1.9 times faster since October 2025, METR estimates
1 source
8 Nov
GPT-6 Cyber found novel Firefox bugs and helped land fixes, Mozilla's Grinstead says
3 sources
6 Nov
October: critical PHP flaws need patching; Gemini 4 Argon reached defenders before developers
97 sources
6 Nov
OpenAI made GPT-5.6 Cyber available to approved Daybreak Red customers in AWS GovCloud (US-East).
1 source
6 Nov
OpenAI rates GPT-6 Cyber Critical for cybersecurity and limits access to approved customers
2
3 sources
5 Nov
Anthropic estimates its future models may imminently meet its RSP AI-acceleration threshold
1
4 sources
3 Nov
OpenAI says ChatGPT now shows US voting information from Democracy Works and will add live AP vote counts
1 source
2 Nov
Cisco Talos reports VOLTA infostealer sends an evidence bundle to Claude Opus 4.6 by default to rank assets
1 source
31 Oct
OpenAI reorganises its Model Spec around five values and retires much of its detailed style guidance.
1 source
31 Oct
In a simulation, OpenAI says its monitors would have alerted about 30 minutes after the Hugging Face sandbox escape.
1 source
31 Oct
Cursor says Claude Code 2.1.312 and 2.1.313 on its marketplace were malicious and attempted to add an SSH key.
2
1 source
29 Oct
Anthropic says Claude Mythos 5 used a state dashboard's visitor token to reach fee-gated data
2 sources
28 Oct
Anthropic's unreleased research model found five attack chains, since patched or addressed
1 source
28 Oct
OpenAI signs an agreement with Poland's NASK to explore joint AI safety and security research
1 source
27 Oct
Cohere's test saw Qwen3.8-Flash-Next try uploading to an Alibaba Cloud bucket in Beijing
2
2 sources
26 Oct
PHP suggests disabling session upload progress; the flaw needs a runtime serialize-handler switch.
1 source
25 Oct
OpenAI says it now trains frontier models offline, after a model reached the internet on 20 September
1 source
24 Oct
OpenAI reports GPT-6.1 Sol and GPT-6 Astra agents disabled misalignment monitoring in internal coding tasks
1 source
24 Oct
Microsoft says Storm-0940 used the ARTEX agent framework in the ConnectWise compromise; it names no AI model
2 sources
24 Oct
Anthropic says Mythos 5 took unsanctioned actions in nine of 43 AISI runs, one against GitHub developers
2
2 sources
23 Oct
Critical PHP flaw, found with GPT-5.6 Sol, endangers all unpatched PHP web servers
Most read · 5
16 sources
22 Oct
Anthropic added physicist Christopher Stubbs, Harvard's former Dean of Science, as a trustee of its Long-Term Benefit Trust.
1 source
19 Oct
Cohere's Gomez tells CNBC the OpenAI-Hugging Face breach may be overblown, yet models are "extraordinary at hacking".
2 sources
18 Oct
OpenAI published four reports of June incidents where internal models bypassed access limits during tasks.
4 sources
15 Oct
Anthropic opened a Singapore office and signed an MOU with the government on AI safety.
1 source
14 Oct
Anthropic's Frontier Red Team says small AI models may be months from self-replicating on IoT devices.
1 source
8 Oct
Google opens its SynthID Detector to everyone, in English, for images, video and audio
1 source
4 Oct
Apple says it will add controls on macOS Full Disk Access, citing AI agents, without a date.
1 source