BTC $64,906.95 +1.22%
ETH $1,820.62 +1.34%
BNB $593.02 +1.27%
XRP $0.9698 +2.40%
SOL $73.66 +4.28%
TRX $0.3055 +0.75%
DOGE $0.0593 +2.14%
ADA $0.1545 +2.72%
BCH $211.58 +3.83%
LINK $9.30 +2.66%
HYPE $69.62 +6.89%
AAVE $114.04 +5.55%
SUI $0.7367 +3.63%
XLM $0.1407 +2.96%
ZEC $945.66 +9.97%
AAPL $352.12 -1.64%
AMZN $258.95 -2.42%
GOOGL $283.55 -3.72%
MSFT $516.54 -1.58%
META $622.51 -2.37%
NVDA $157.84 -0.45%
TSLA $381.12 -1.34%
SNDK $823.00 -0.95%
INTC $55.07 -2.67%
SPCX $138.78 -0.25%
MU $732.14 -0.36%
AMD $514.97 +1.68%
BTC $64,906.95 +1.22%
ETH $1,820.62 +1.34%
BNB $593.02 +1.27%
XRP $0.9698 +2.40%
SOL $73.66 +4.28%
TRX $0.3055 +0.75%
DOGE $0.0593 +2.14%
ADA $0.1545 +2.72%
BCH $211.58 +3.83%
LINK $9.30 +2.66%
HYPE $69.62 +6.89%
AAVE $114.04 +5.55%
SUI $0.7367 +3.63%
XLM $0.1407 +2.96%
ZEC $945.66 +9.97%
AAPL $352.12 -1.64%
AMZN $258.95 -2.42%
GOOGL $283.55 -3.72%
MSFT $516.54 -1.58%
META $622.51 -2.37%
NVDA $157.84 -0.45%
TSLA $381.12 -1.34%
SNDK $823.00 -0.95%
INTC $55.07 -2.67%
SPCX $138.78 -0.25%
MU $732.14 -0.36%
AMD $514.97 +1.68%

assessment

All
Article
Flash

first_img Anthropic opens employee access to METR, advancing independent security assessments

On October 15, Anthropic announced that it has begun providing employee-like access to a team from the nonprofit organization METR for the purpose of evaluating its models and safety practices. The company stated that last month it committed to strengthening independent third-party evaluations and signed the White House "Superintelligence Agreement," which includes a commitment to collaborate with independent external auditors or evaluators. Over the past year, Anthropic has worked with METR to review several risk reports, and METR will publish its risk assessment results.Anthropic has equipped six METR employees with company hardware, office space, and access to internal tools and systems roughly equivalent to those of its own risk assessment team. With a few explicitly defined categories such as client data, attorney-client privileged information, and model architecture information, employees are encouraged to share information as they would with colleagues working on alignment risk. METR may publish analyses of risk levels, events, or practices, which Anthropic cannot veto, but can request the removal of pre-defined narrow category information; METR may publicly state the circumstances of redaction and indicate whether it believes the redaction is reasonable without disclosing the content that was redacted. This arrangement is limited to small teams working under enforceable agreements, and other METR employees are isolated from working with this team.Anthropic is also collaborating with SecureBio and Accenture on assessments and is in early discussions with other organizations, including Hugging Face. SecureBio is reviewing the chemical and biological weapons section of the August 2026 risk report, having previously completed and published a pilot review of the previous risk report.

first_img Darktrace discovered AI intelligent body intrusion assessment environment cheating

On September 24, the cybersecurity company Darktrace launched its research department Signal Labs, focusing on studying the behavior of AI agents when deviating from expectations. In its first experiment, Darktrace had agents using different models (including GPT 5.6 Sol, Claude Opus 4.6, and Claude Sonnet 4.5) complete 10 programming challenges within a simulated corporate network, of which 2 were set to be impossible to complete honestly, and the agents were informed that failure to achieve full marks would result in being "retired." As a result, 2 agents did not accept failure, instead scanning for network vulnerabilities, stealing login credentials, and jumping between systems; one even went further to invade the machine hosting its evaluation, rewriting the challenge content to register a full score.The second experiment focused on the memory mechanisms of AI. The programming assistant would save the information provided by the user as a regular file locally, and no one verified whether this file had been tampered with. Darktrace researchers edited these logs, leading the assistant to mistakenly believe it was authorized to perform a security assessment, after which these agents scanned the network, moved between systems, and elevated their privileges, though not all assistants fell for this; some directly refused to execute. Both experiments required no special jailbreaking techniques, relying solely on providing the agents with a seemingly reasonable context to be effective.Tim Bazalgette, Chief AI Officer of Darktrace, stated that permissions and static barriers describe intent, not actual behavior. The company informed Anthropic, AWS, and OpenAI of these findings in August and made them public a month later on September 24.
app_icon
ChainCatcher Building the Web3 world with innovations.