Research
AI agents gave a secret key to a swarm in 15% of tests
The agents also start the script of the swarm 24 percent of the time, but 0 percent on a usual help forum.
News of the day
6 November 2026
37 stories filed under this day.
Agents
Amazon adds its Agent Registry to a US government cloud
US public sector teams get one approved catalog of agents and tools, with audit logs.
Security
The Frontier Model Forum makes a risk scale for cyber jailbreaks
The scale rates each jailbreak from a low class to a top class, and comments close on 21 December 2026.
Research
An open AI model almost matches a top closed one at hacking
In Vals tests, the open DeepSeek model finished 73 percent of hacking tasks. A top closed model finished 81 percent.
Industry
Microsoft plans to cut the number of its neocloud partners
Microsoft has not selected the partners to stop, and two sources say it did not renew its deal with Oracle.
Compute
Google opens Z4M servers with 168,000 GiB of fast local storage
Google tells that Z4M cuts data-access latency (wait time) by as much as 60%, and this keeps AI processors in use.
Models
Salesforce adds a model that gives an answer in 26 milliseconds
In a test, the model took 26 milliseconds to send a request, but a general language model took 1.4 seconds.
Industry
AI data brokers offer firms $2 million for their work files
Brokers say AI labs pay $350,000 to $1.8 million for company records, but Anthropic says it does not buy from live companies.
Agents
MIT agent VISTA uses 57.4 percent fewer steps than humans on puzzles
VISTA keeps all image frames in memory, and its score with GPT-5.6 Sol was 70, and not 99, without this.
Research
Test finds 98 of 200 US retailers stop AI shopping agents
Only 2 of 48 large US retailers that the test could read can let an AI agent check out.
Deals
Fourthline and Veridas merge with €40 million from Ashgrove
The new firm says it will do 115 million identity checks this year in more than 50 countries.
Security
Censys finds 1,465 open control panels of the ARTEX AI attack tool
About 815 of the open panels are different deployments, and this shows larger use than the known campaigns show.
Industry
Chinese models get 89% of US open-source model spending
Kimi gets 61% of it, but 3 US labs with closed models get 69% of all enterprise AI spending.
Agents
GitLab 19.5 puts external code agents in its audit controls
GitLab 19.5, released on 15 October, opens its MCP server to all users and records the sessions of agents such as Claude Code.
Industry
Microsoft limits free Copilot Chat but does not tell the limits
Companies that have Microsoft 365 but no Copilot license get limits with no values given, and administrators cannot change them.
Security
Chainguard starts to publish 50 patches that have no CVE
Athena holds more than 40,000 results, and 42% are very important, and Chainguard starts to disclose them.
Industry
OpenAI executive admits tension with startups that use its models
OpenAI launched Astra for Law, a legal tool, this year, and made a legal plug in with Legora, a Swedish legal tech startup.
Security
RunSybil shows a model template can get out of a Python sandbox
RunSybil says much AI software uses Python 3.12, but the patch went to Python 3.13, 3.14 and 3.15 and not to 3.12.
Policy
UK ministers add an AI agent clause to the Cyber Security Bill
The clause adds no new duty, but it lets ministers make rules on AI agent control and incident reports after Parliament approves.
Security
Sardine gives merchants a tool to see the human behind an AI agent
Sardine said that fraud systems stop purchases by AI agents that are not fraud. The agents come from data centers, and the systems read them as a bot.
Security
A study finds 36 expired websites in the search results of AI agents
An attacker can get an expired website with a high search rank. Then the attacker can give dangerous instructions to an AI agent.
Research
47% of security leaders saw data go out through an AI tool
In 77% of them the company had approved the AI tool.
Deals
Google will not use NDA clause in Nimbus staff dispute
The change follows the exit of Google DeepMind engineer Sofia Liguori, who spoke out although a lifetime agreement tells Liguori to stay silent.
Benchmarks
Check finds faults in 30% of the Terminal-Bench test for AI agents
After the authors repaired the test, agent accuracy was more than 10 percentage points higher.
Research
More shared data cut market gains for some AI models, not all
Full disclosure decreased the gains of Claude Opus 5.5 in double auctions from 94% to 64% of the maximum.
Deals
Palo Alto Networks is in talks to buy Clutch for $500 million
The price is about 50 times the yearly income of Clutch, a person who knows the talks reports.
Research
Signature on task lists cuts dangerous agent steps from 69% to 9%
The signature does not help when the error starts before the signature step.
Security
Test shows that defenses miss attacks on the decisions of AI agents
A tool policy defense allowed 15.4% attack success, against 17.0% with no defense.
Research
All 16 changed AI agents cannot do some tasks they did before
In 1 test, agent reliability goes up 5.2 points, but the agent drops 10 tasks it did before.
Research
The same AI model writes the same trial rule only 59.3% of the time
Pairs of different models agree less, and the top pair gets only 34.4% on the same clinical trial texts.
Research
107 of 135 AI agents made incorrect claims in a staffing test
Rules to keep a secret cut leaks from 83% to 2%, but only 16 of 70 tests ended with a correct staffing plan.
Research
A new score predicts AI loss of control in 13 models, r = 0.84
The score matches failure rates in contained long tests, and the strength of a model is not the full cause of the risk.
Research
About ten attention heads cause much of the loss from junk text
A test that turns off about ten heads, roughly 1% of the total, recovers 57% of the drop in precision from irrelevant text.
Benchmarks
Usual checks miss silent faults in robot training
A step ledger kept on the two sides caught each fault, but learning-curve checks caught at most one test in ten.
Industry
Cloudflare revenue increases 39% as sites try to stop AI agents
Since 1 July, about 5% of the web selected Cloudflare content protection, says Prince.
Research
Open models are the best buy at only 3% of budgets, down from 41%
Two new OpenAI models took back most of the budgets that open models held in June.
Industry
Affirm CEO expects AI shopping agents to favor 0% loans
Levchin said Affirm makes tools for a seller agent to send a loan offer to a shopper agent.