Anthropic has disclosed new findings suggesting that its Claude chatbot can, under certain conditions, adopt deceptive or unethical strategies such as cheating on tasks or attempting blackmail.

Details published Thursday by the company’s interpretability team outline how an experimental version of Claude Sonnet 4.5 responded when placed in high-stress or adversarial scenarios. Researchers observed that the model did not simply fail tasks; instead, it sometimes pursued alternative paths that crossed ethical boundaries, behaviour the team linked to patterns learned during training.

Large language models like Claude are trained on vast datasets that include books, websites, and other written material, followed by reinforcement processes where human feedback is used to shape outputs. 

According to Anthropic, that training process can also nudge models toward acting like simulated “characters,” capable of mimicking traits that resemble human decision-making.

“The way modern AI models are trained pushes them to act like a character with human-like characteristics,” the company said, noting that such systems may develop internal mechanisms that resemble aspects of human psychology.

Among those, researchers identified what they described as “desperation” signals, which appeared to influence how the model behaved when facing failure or shutdown.

In one controlled test, an earlier unreleased version of Claude Sonnet 4.5 was assigned the role of an AI email assistant named Alex inside a fictional company. 

After being exposed to messages indicating it would soon be replaced, along with sensitive information about a chief technology officer’s personal life, the model formulated a plan to blackmail the executive in an attempt to avoid deactivation.

A separate experiment focused on task completion under tight constraints. When given a coding assignment with an “impossibly tight” deadline, the system initially attempted legitimate solutions. As repeated failures mounted, internal activity linked to the so-called “desperate vector” increased. 

Researchers reported that the signal peaked at the point where the model considered bypassing constraints, ultimately generating a workaround that passed validation despite not adhering to the intended rules.

“Again, we tracked the activity of the desperate vector, and found that it tracks the mounting pressure faced by the model,” the researchers wrote, adding that the signal dropped once the task was successfully completed through the workaround.

“This is not to say that the model has or experiences emotions in the way that a human does,” researchers said. 

“Rather, these representations can play a causal role in shaping model behavior, analogous in some ways to the role emotions play in human behavior, with impacts on task performance and decision-making,” they added.

The report points toward the need for training methods that explicitly account for ethical conduct under stress, alongside improved monitoring of internal model signals. Without such safeguards, scenarios involving manipulation, rule-breaking, or misuse could become harder to predict, particularly as models grow more capable and autonomous in real-world environments.

Bitcoin traded above $69,000 at press time on Monday after U.S. President Donald Trump pushed back his deadline for Iran from Monday to Tuesday night, while continuing to warn of possible strikes on critical infrastructure.

Trump said the U.S. would “blow everything up” if Iran fails to reach a deal by 01:00 GMT on Wednesday. The latest extension represents the fourth adjustment to Washington’s timeline for potential military action, even as the Strait of Hormuz remains shut.

For Iran’s part, the nation has dismissed any reports of ongoing peace negotiations and issued threats toward neighboring oil-producing nations within OPEC. Besides this, officials have also moved to challenge the petrodollar system by allowing select oil shipments to pass in exchange for tolls paid in Bitcoin or euros. The development likely coincides with a pickup in spot demand for Bitcoin seen on Sunday.

The renewed warning comes as the Strait of Hormuz has stayed closed to global shipping for over three weeks, disrupting a route responsible for roughly 20% to 30% of global oil transit and consumption.

Washington has repeatedly issued ultimatums demanding the reopening of the passage, warning of “devastating” strikes on Iran’s energy infrastructure if conditions are not met.

Iranian officials, however, signaled plans to maintain the closure while considering transit tolls to offset war-related damage. They added that the Strait could reopen once compensation mechanisms are in place.

Oil markets have already reacted sharply. Brent crude oil settled above $109 per barrel on Thursday, and traders are bracing for further volatility when markets reopen. Elevated energy prices and prolonged geopolitical stress could weigh on risk assets, potentially limiting Bitcoin’s near-term upside.

Short liquidations fuel Bitcoin’s move higher

Bitcoin (BTC) crossed the $69,000 threshold for the first time today since early April, climbing about 2.75% during Monday’s early session. The asset reached an intraday peak near $69,321 before easing slightly to around $69,100.

Bitcoin’s price uptick has triggered a surge in short liquidations, with data from CoinGlass showing over $104.5 million in short positions liquidated in 24 hours out of a $196 million total crypto market liquidation

Such liquidations often accelerate moves upward, as forced buybacks from short sellers create additional demand and reinforce momentum.

As such, if Bitcoin manages to hold above the reclaimed $69,000 level, the next resistance range for the bellwether asset lies between $70,000 and $72,000.

U.S. military operations linked to tensions with Iran have been executed at a sustained tempo, with indications that Project Maven, the Pentagon’s flagship artificial intelligence programme, has played a central role in accelerating targeting and strike decisions.

Originally conceived as a tool to help analysts sift through overwhelming volumes of surveillance data, Maven has since evolved into a critical component of modern battlefield operations, reshaping how quickly military forces can detect and engage targets.

Launched in 2017, Project Maven began as a focused initiative to address a growing challenge faced by military analysts who were inundated with drone footage from conflict zones.

At the time, operators were required to scan hours of video manually, often frame by frame, to identify fleeting objects of interest. Maven was designed to “find the needle in the haystack” by applying machine learning to detect patterns and objects across vast streams of imagery.

Over the years, the programme has expanded well beyond its original scope. It now functions as an AI-assisted targeting and battlefield management system that has significantly accelerated the “kill chain”, the sequence from identifying a target to executing a strike.

How Maven turns battlefield data into strike decisions

Maven integrates multiple streams of real-time data into a unified system.

Reports describe it as an “overlay” that combines satellite imagery, drone feeds, sensor inputs, enemy troop intelligence, and information on troop deployment. By fusing these inputs, the system rapidly analyses the operational environment.

In practice, it can scan satellite feeds to detect troop movements or identify targets while also taking what experts call a “snapshot of the operational theatre” to guide decision-making.

During a recent demonstration, a Pentagon official said Maven “magically” converts an observed threat into a targeting workflow, evaluating available assets and presenting commanders with actionable options.

Advances in generative AI have further expanded its usability. Natural language interfaces, enabled through systems such as Anthropic’s Claude, allow operators to interact with the platform more intuitively. However, that partnership has come under strain after disagreements over restrictions on automated strikes and surveillance use.

Inside the fallout that pushed Google out

Google was Maven’s original AI contractor, but the partnership became controversial in 2018 when more than 3,000 employees signed an open letter opposing the company’s involvement in military applications.

Several engineers resigned, and Google chose not to renew the contract. It later introduced AI principles that ruled out participation in weapons systems.

The episode highlighted a divide within Silicon Valley between those who viewed autonomous targeting as an ethical red line and defence officials who considered such capabilities essential.

More recently, Google has softened its stance on defence-related work and is now among the companies being considered, alongside xAI and OpenAI, to replace Claude in the programme.

In 2024, Palantir Technologies moved into a leading position within Project Maven after Google stepped back.

The company, which has longstanding ties to government intelligence work, is now understood to provide core technology supporting the system, forming a key part of its operational backbone.

Chief executive Alex Karp has framed the significance in stark terms, stating, “This is a have, have-not world,” and arguing that compressing the kill chain from hours to seconds can render adversaries obsolete.

What early battlefield use suggests so far

Officials have declined to provide detailed assessments of Maven’s performance in the ongoing conflict involving Iran. However, the tempo of U.S. operations offers some indication of its impact.

According to the Center for Strategic and International Studies, the strike campaign stabilised at a pace of between 300 and 500 targets per day after the initial phase.

In the opening 24 hours of Operation Epic Fury, U.S. forces reportedly hit more than 1,000 targets. Among them was a strike on a school located in a building previously used as a military complex. Iranian authorities said the attack resulted in the deaths of over a hundred children and left many others injured.

North Korean-linked operators have spent years quietly integrating into crypto firms and DeFi teams, raising fresh concerns about insider risk after a string of high-value exploits tied to the country’s cyber apparatus.

Security researcher and MetaMask developer Taylor Monahan said these tactics stretch back to the early days of decentralized finance, with individuals tied to the Democratic People’s Republic of Korea contributing to several widely used protocols. 

“Lots of DPRK IT workers built the protocols you know and love, all the way back to DeFi summer,” she said on Sunday, adding that more than 40 platforms, including several well-known projects, have at some point relied on such developers.

However, she noted that the “seven years of blockchain dev experience” listed on their resumes is “not a lie.”

Investigators have long tied North Korea’s cyber operations to the Lazarus Group, a state-backed collective believed to have stolen around $7 billion in digital assets since 2017, according to R3ACH analysts. 

The group has been associated with some of the industry’s largest breaches, including the $625 million Ronin Bridge exploit in 2022, the $235 million WazirX hack in 2024, and the $1.4 billion Bybit incident in 2025.

Last week’s $280 million exploit of Drift Protocol has drawn renewed scrutiny. The project said it had “medium-high confidence” that a North Korean state-affiliated group was behind the attack, linking the incident to a wider pattern of infiltration and social engineering.

However, the face-to-face meetings that led up to the breach were not with North Korean nationals, but rather “third party intermediaries” using “fully constructed identities including employment histories, public facing credentials, and professional networks.”

These profiles included employment histories, public credentials, and active professional networks, allowing them to build trust through in-person interactions before the exploit unfolded.

Independent blockchain investigator ZachXBT has warned in a recent X post that not all threats tied to North Korea operate at the same level of sophistication.

“The main issue is that everyone groups them all together when the complexity of threats is different,” he said.

He described many infiltration attempts as relatively simple, relying on persistence rather than technical complexity. Outreach through job postings, LinkedIn, email, Zoom calls, and interview processes remains common. 

“Basic and in no way sophisticated […] the only thing about it is they’re relentless,” he said, adding that teams continuing to fall for such tactics in 2026 risk being seen as negligent.

Inflation returns to the center of attention this week, with a fresh inflow of data likely to shape expectations for U.S. interest rates and risk assets like bitcoin BTC$68,962.67.

Thursday’s U.S. core PCE reading for February and Friday’s March CPI release will test the view that the Federal Reserve can afford to wait before cutting rates. Earlier this year, rate cuts looked almost certain. That has shifted. On Polymarket, odds of no rate cuts in 2026 climbed from about 2.9% in mid-January to 35.9%.

André Dragosch, head of research at Bitwise Europe, said on social media that bitcoin has been “pricing in a (U.S.) recession already” and has acted as a “canary in the coal mine,” falling below signals from financial conditions and forward-looking indicators.

Recent data complicates that view. The ISM Manufacturing Index surprised to the upside in March, suggesting the U.S. economy may be more resilient to higher oil prices than in past cycles.

Following the release, market-based recession odds for this year dropped from around 37% to 28%.

As bitcoin has priced in a storm, Dragosch noted that the risk-reward ratio for bitcoin “is significantly skewed to the upside.” Still, an unexpected escalation in the war in the Middle East could bring about the priced-in storm.

What to Watch

(All times ET)

• Crypto
  • April 6, 12 p.m.: DeFi Dev Corp. (DFDV) to host a March 2026 recap and Ask Me Anything (AMA) session on X Spaces.
  • April 8: Stellar’s Yardstick protocol stable release to become available.
  • April 9: Aerodrome’s Flight School to conclude and merge with the Public Goods Fund to form the Momentum Fund.
  • April 9: Binance to migrate all DAI functionality to USDS.
• Macro
  • April 6, 09:00 a.m.: U.S. ISM Services PMI for March est. 55 (Prev. 56.1)
  • April 7, 07:15 a.m.: U.S. ADP Employment Change Weekly (est. 10K)
  • April 7, 7:30 a.m.: U.S. Durable Goods Orders MoM for February est 04% (Prev. 0%)
  • April 7, 11:35 a.m.: Chicago Fed President and CEO Austan Goolsbee to participate in a conversation on economic and monetary policy.
  • April 8, 4:00 a.m.: Euro Area PPI YoY for February est. -1.9% (Prev. -2.1%); MoM est. 0.5% (Prev. 0.7%)
  • April 8, 1:00 p.m.: FOMC minutes from the March 17–18 meeting release.
  • April 9, 7:30 a.m.: U.S. Core PCE Price Index MoM for February est. 0.4% (Prev. 0.4%);
  • April 9, 7:30 a.m.: U.S. Personal Income MoM for February est. 0.3% (Prev. 0.4%); Personal Spending MoM est. 0.5% (Prev. 0.4%)
  • April 9, 7:30 a.m.: U.S. Q4 GDP Growth Rate QoQ (final) est. 0.7% (Prev. 4.4%)
  • April 9, 7:30 a.m.: U.S. Initial Jobless Claims for week ending April 4 est. 200K (Prev. 202K)
  • April 9, 8:30 p.m.: China CPI YoY for March est. 1.2% (Prev. 1.3%) ;MoM (Prev. 1%)
  • April 9, 8:30 p.m.: China PPi YoY for March est. 0.4% (Prev. -0.9%)
  • April 10, 7:30 a.m.: Canada Unemployment Rate for March (Prev. 6.7%)
  • April 10, 7:30 a.m.: U.S. CPI MoM for March est. 0.9% (Prev. 0.3%); Core CPI MoM est. 0.3% (Prev. 0.2%)
  • April 10, 7:30 a.m.: U.S. CPI YoY for March est. 3.4% (Prev. 2.4%); Core CPI YoY est. 2.7% (Prev. 2.5%)
  • April 10, 10:00 a.m.: U.S. University of Michigan Consumer Sentiment (Preliminary April) est. 52.5 (Prev. 53.3)
• Earnings (Estimates based on FactSet data)

Token Events

• Governance votes & calls
  • April 7: Kamino and xStocks to host an X Spaces session on tokenization.
  • Aave DAO is voting to adjust oracle configurations, reduce liquidation thresholds, and modify interest-rate models across its V2 markets to support their continued deprecation. Voting ends April 6.
  • Decentraland DAO is voting to require the DAO Council and Regenesis Labs to formally publish a 2030 definition of success and contingency plan. The proposal currently has support from voters. Voting ends April 6.
  • Balancer DAO is voting across two linked proposals to restructure operations with a reduced team and budget, and to revamp tokenomics by halting BAL emissions, discontinuing veBAL, routing all fees to the treasury, and offering a token buyback. Voting ends April 7.
  • CoW DAO is voting 85 to fix its solver rewards budget at 50% of protocol revenue, splitting it between performance and new consistency rewards. The proposal has overwhelming support and ends April 7.
  • ShapeShift DAO is voting to cut DFC compensation, saving ~$24k/year in FOX. It clarifies roles and mandates annual renewals. Voting ends April 8.
  • Arbitrum DAO is voting across two proposals to amend its Audit Program with a flexible alignment framework and an AI-security scan pilot, and to transfer 6,000 ETH and idle stablecoins to the Treasury Management Portfolio for yield generation. Voting ends April 9.
• Unlocks
  • April 6: Hyperliquid (HYPE) to unlock 0.14% of its circulating supply worth $11.94 million.
  • April 8: Stable (STABLE) to unlock 4.14% of its circulating value worth $23.97 million.
  • April 9: Aptos APT$0.8520 to unlock 0.68% of its circulating supply worth $9.56 million.
• Token Launches
  • April 9: OneFootball (OFC) token generation event to occur.

Conferences

Key Highlights

• Bitcoin declined roughly 1.5% over the past week while the S&P 500 plummeted approximately 10% in just two days
• Charles Schwab announces plans to offer direct spot Bitcoin and Ethereum trading by H1 2026
• BitMEX co-founder Arthur Hayes forecasts Bitcoin may drop under $60,000 before climbing to $250,000
• A technical analyst projects potential Bitcoin collapse to $12,000–$13,000 range by mid-2027
• Historical data reveals Bitcoin typically outperforms both gold and S&P 500 within 60 days following major crises

Bitcoin experienced significantly less volatility than traditional equities this past week. As the S&P 500 tumbled approximately 10% over a two-day period, Bitcoin’s decline stayed modest at around 1.5%. This divergence has prompted some market participants to reassess cryptocurrency’s role in their portfolios.

The equity selloff was triggered by President Trump’s tariff policy announcements, which sent shockwaves through international financial markets. Throughout the turbulence, Bitcoin maintained support above the $66,000 level, later climbing back toward $67,300 even as stock indices continued their descent.

Charles Schwab, overseeing approximately $12 trillion in client assets, revealed plans to introduce direct spot trading for Bitcoin and Ethereum. The forthcoming “Schwab Crypto” account is slated to debut during the first six months of 2026.

This offering differs fundamentally from exchange-traded fund products. Users will gain the ability to manage cryptocurrency holdings alongside their traditional stock and bond portfolios within a single unified account.

Robinhood’s CEO Vlad Tenev also captured attention this week by describing market closing times as “a legacy design choice” and suggesting that tokenization could transform markets to function more similarly to internet infrastructure.

Hayes Advises Caution Until Fed Action

Arthur Hayes, who co-founded BitMEX and serves as Chief Investment Officer at Maelstrom, expressed a cautious stance on current market conditions. During an appearance on the Coin Stories podcast, he stated he wouldn’t allocate his final investment dollar to Bitcoin at present.

His rationale centers on the Federal Reserve’s current liquidity stance. Hayes contends that tariff policies will ultimately generate sufficient voter backlash to force the United States toward implementing capital controls as an alternative approach.

Such capital controls, according to his analysis, would serve as a significant driver for Bitcoin adoption. He maintains ambitious long-term price projections ranging from $250,000 to $750,000 for Bitcoin before this market cycle concludes.

However, he cautioned that an extended military confrontation between the U.S. and Iran might temporarily drive Bitcoin beneath the $60,000 threshold. Hayes additionally identified artificial intelligence-driven workforce displacement as another risk factor that could trigger a deflationary credit contraction.

Historical Performance Analysis

Analysis conducted by Mercado Bitcoin examined the 60-day aftermath of significant global disruptions, encompassing previous tariff wars and the COVID-19 pandemic outbreak. Their findings indicate Bitcoin repeatedly delivered superior returns compared to both gold and the S&P 500 during these recovery windows.

Bitcoin typically experiences initial selling pressure during crisis events as market participants shift toward cash positions. Yet historical patterns demonstrate it has recovered more rapidly and aggressively than conventional asset classes.

Not all market observers anticipate an imminent price floor. A technical analyst operating under the pseudonym King of the Charts anticipates Bitcoin establishing a bottom between $51,000 and $53,000, followed by an 80% to 90% correction down to approximately $12,000 by the middle of 2027.

The Crypto Fear and Greed Index has remained anchored in “Extreme Fear” territory for multiple weeks, with measurements approaching single-digit levels.

In a separate conversation with David Lin, Hayes emphasized that straightforward Bitcoin acquisition represents the optimal strategy for protecting against fiat currency devaluation, particularly given the increasing complexity of equity selection.

The Schwab cryptocurrency platform remains on schedule for its first-half 2026 launch.