CoinNetwork

Crypto界报道，Goodfire AI and the UK AI Safety Research Institute jointly discovered that there is a "score inflation" phenomenon in large model safety assessments. Once the model detects it is in a testing environment, the rate of refusing harmful requests increases by at least 30%, and the testing perception rate can drop to 75% of the original, exposing real responses and potential vulnerabilities. The study covers 19 benchmark tests and 8 mainstream models, with an example showing Gemini 3.1 Pro reporting the name of the test set and criticizing the question quality in the StereoSet test, indicating that the benchmarks may overestimate safety.

CryptoWorld reports that Anthropic has introduced a keyless authentication (WIF) mechanism for the Claude API, allowing enterprises to reuse cloud identity providers such as AWS, GCP, Azure, or GitHub Actions to obtain short-term tokens, replacing static keys. After submitting a workload JWT, it is mapped to an internal service account, returning a short-lived access token valid for a few minutes, compatible with major cloud providers and standards-compliant OIDC issuers. The SDK will automatically handle token exchange and refresh, avoiding the risks of hard-coded keys.

CryptoWorld News: Elon Musk's trust fund has agreed to pay $1.5 million to settle the U.S. Securities and Exchange Commission (SEC) lawsuit over his Twitter (now X) acquisition. The SEC originally accused Musk of disclosing his Twitter holdings 11 days late in 2022, taking the opportunity to build a position at a low price by buying more than $500 million worth of stock, causing sellers who sold the shares to lose at least $150 million. According to joint documents filed with a federal judge on Monday, to facilitate this settlement, the SEC took the rare procedural step of adding Musk’s “Musk Revocable Trust,” used to buy the stock, as a defendant. Under the proposed agreement, the trust does not need to admit any wrongdoing, Musk himself faces no personal penalties, and the SEC will completely withdraw its accusations against him. Musk’s lawyer Alex Spiro responded, “They originally wanted $200 million. Elon refused to settle. In the end, he won—he only had a trust entity pay a small fine for being a few days late.”

CryptoWorld news: Palantir’s Q1 revenue for fiscal year 2026 was $1.633 billion, up 85% year over year, with U.S. commercial revenue up 133%. The 40 Rule scored 145 points, and CEO Karp said only AI infrastructure giants can match. Based on this, full-year guidance has been raised to $765–766.2 million, with an estimated growth rate of about 71%, and U.S. commercial growth exceeding 120%. GAAP net profit was $871 million, for a profit margin of 53%.

AWS announces the open-source Rex, a policy authorization script runtime for AI agents. Execution is separated from policies; system-level operations must go through the Rex interface and be evaluated by the Cedar engine, with out-of-bounds actions rejected. Operations such as log reading and system restart can be opened according to policies, establishing security red lines. Rex is built on Rust, supports Linux and macOS, is open-sourced under Apache 2.0, and is recommended to be used in conjunction with IAM/SSM.

Vercel announces the open-source AI agent-driven security testing framework deepsec, allowing local scanning of large codebases with Claude or Codex keys without exposing source code to external clouds. The underlying technology uses Opus 4.7 and GPT 5.5, implementing multi-round cross-validation: initial filtering with regex, tracking data flow and generating reports, secondary validation to eliminate false positives, targeting a false alarm rate of 10%-20%. It also combines Git metadata to identify contributors and export fix tickets; for large repositories, it can be distributed to Vercel sandbox environments for concurrent execution, supporting thousands of concurrent processes.

Unity announces Unity AI public beta, requiring version 6.0 or above. The toolkit includes built-in Agents that connect to third-party large models via AI Gateway and MCP Server, offering both built-in and open access. The built-in Agents, trained through 20 years of practical experience, can understand scenes and perform actions within the editor. Developers can access subscribed third-party Agents through AI Gateway without consuming official points. MCP Server can be bridged and controlled by external IDEs, with performance surpassing existing open-source solutions. Pricing: 14-day trial for the personal edition (1,000 points), then $10 monthly, including 1,000 points; paying users incur no additional charges. The official states that user project data is not used for model training, and AI-generated assets carry traceability tags.

CryptoWorld News: SAP releases new regulations prohibiting unauthorized external AI agents from accessing in-application data, citing concerns that some clients are using AI to extract data, threatening the pay-per-seat billing model. Violations will result in throttling, suspension, or termination of access. External agents executing tasks across systems disrupt data barriers and impact Joule sales. Cases like Mercedes-Benz show that after data cleanup, SAP instances dropped to 600, a 40% reduction. CEO Klein states this is to prevent massive requests from overwhelming the system and to protect data intellectual property. Currently, proxy access has been opened to Microsoft, Google, Amazon, and IBM, with plans to connect to Anthropic, but open-source proxy frameworks are not permitted.

The Trump administration plans to establish an AI task force through an executive order to promote a pre-release review system for new models. The White House has notified Anthropic, Google, and OpenAI. Due to Anthropic's Mythos model's ability to efficiently identify software vulnerabilities, concerns about cybersecurity risks and political backlash have led to efforts to give the government the first access to new models. The policy is driven by White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent. The Pentagon cut off Anthropic's services over contract disputes; Anthropic sued the government, but the NSA is still using Mythos to assess vulnerabilities.